Field Notes

meta

Log

StatusStableUpdatedSupportsnotes

2026-07-04

  • Created the initial filesystem wiki scaffold.

2026-07-09

  • Clarified semantic-line-break guidance in AGENTS.md and manually corrected bad SemBr throughout src/ to keep tightly bound grammar and ideas on the same source line.
  • Updated the agent workflow so assistants answer from src/, research beyond the local wiki when useful, preserve linked sources under src/sources/, and synthesize durable findings back into the wiki.
  • Made src/ the complete editor vault and moved immutable source artifacts into src/sources/.
  • Added wanted-page redlinks so agents can mark future articles without creating ungrounded stubs.
  • Added commit-message AI provenance trailers for tracking which harness and model made wiki edits.
  • Added Privacy business thesis as an initial synthesis for a possible privacy, security, antisurveillance, and anticensorship business.
  • Added EU and Swedish data removal rights to distinguish GDPR erasure, search-result delisting, and the Swedish utgivningsbevis complication for a data-rights service.
  • Added Integritetsskyddsmyndigheten as the regulator and complaint-escalation path relevant to those requests.
  • Preserved local copies of IVPN Unlinked Access and Obscura trust 2-party relays and QUIC, then added Unlinked subscription architecture to capture the business idea of one bill without one cross-service identity.

2026-07-10

2026-07-12

  • Added four cross-cutting concepts from the expanded privacy synthesis: Data minimization, Data as coercive power, High-consequence data, and Confidentiality as safety infrastructure. Linked them into the case, product architecture, governance, customer, lawful-extraction, and digital-safety-service notes so case studies now lead to reusable design and market concepts.
  • Imported two ChatGPT Deep Research reports as matched Markdown/PDF pairs: Privacy, security, and the ethics of resisting lawful extraction and Justifications for caring about privacy rights. The source cards explain that the Markdown is searchable but has broken internal citation tokens, while the PDFs preserve readable numbered endnotes and source URLs.
  • Absorbed the reports’ substantive argument into Case for privacy and security, including the rights framework, harm model, ten case studies, lawful-extraction analysis, proportionality test, objections, market implications, and operating principles. Added the intact PDF artifacts and underlying endnote sources to the synthesis metadata so its evidence is visible without routing through the source cards.
  • Added Ethics of resisting lawful extraction to distinguish legal authorization from ethical justification through necessity, proportionality, third-party harm, institutional reliability, and irreversible disclosure. Updated Case for privacy and security, Privacy customer segments, and Privacy business thesis with the reports’ data-to-leverage model, incident-grounded customer needs, and harm-specific commercial positioning.
  • Added Mermaid diagrams to six notes whose structure was already diagrammatic in prose but had no visual: the trust ladder in Private AI trust boundaries, the request flow in Confidential AI computing, the route selector in Private AI routing, the customer path in Post-seizure digital recovery, the green-amber-red gates in Private AI legal roles, and the staged plan in Privacy venture roadmap. Each keeps navigable wikilinks in nearby prose, uses plain note-title labels inside the fence, and avoids smart typography so lint stays clean.
  • Refined the no-anti-forensics boundary in Post-seizure digital recovery so it forbids wiping a device under seizure, preservation order, or evidentiary custodial chain, rather than condemning remote wipe as such. Standard anti-theft remote wipe of a device the customer still owns — as Google Find My Device and Apple Find My provide — is a legitimate security feature and was being conflated with evidence destruction.
  • Fixed the desktop left sidebar so it stays fixed while scrolling, matching the right table-of-contents pane. bin/wiki.py now pins .sidebar with position:fixed and a max() calc that tracks the centered grid, while main and .toc get explicit grid-column placement so the reserved sidebar column remains empty.
  • Made just build skip the static rebuild when nothing in src/, bin/wiki.py, or share/ is newer than a dist/.build-stamp written at the end of each build. cmd_build gained a --force flag and just rebuild calls it, so agents can force a full regenerate when the stamp looks stale. Updated AGENTS.md so agents run just build after any wiki edit — it is a cheap no-op when nothing changed — and know about just rebuild for forced regeneration.
  • Made the desktop table-of-contents pane scroll independently when a note has many sections, keeping its page tools, including the theme button, reachable at the end of the pane.
  • Added Tag map, an auto-generated companion to Knowledge graph map that renders a tag co-occurrence graph from frontmatter tags. Every tag is a sized circle (size scales with note count, so the map doubles as a frequency-weighted tag cloud), and edges connect tags that share at least two notes, thickening with co-occurrence count. bin/wiki.py gained tag_graph_mermaid and a tag-graph fence expander that mirrors the existing wiki-graph mechanism, reusing the Mermaid ELK layout and interactive toolbar without the note-graph edge bundling. Wired the new page into Index and the then-existing “Ontology” page.
  • Normalized tags to mid-sentence natural case and recorded the convention in AGENTS.md: lowercase common nouns and adjectives, natural capitalization for proper nouns and acronyms, and a hyphen only where English would hyphenate the phrase mid-sentence (local-first), leaving the rest as spaced phrases (confidential computing, lawful access, data governance, digital forensics) so each tag matches the prose surface form. The map had surfaced split variants: confidential computing and vpn on three notes alongside confidential-computing and VPN. Normalized the hyphenated common-noun tags to their spaced forms across 12 notes — confidential computing, lawful access, data governance, digital forensics, civil society, and case studies — and the lowercase vpn to VPN on Obscura trust 2-party relays and QUIC, since VPN keeps natural acronym capitalization mid-sentence. The distinct tag count stayed at 50 and the edges merge cleanly on the graph. Updated the Tag map legend to describe the convention.
  • Consolidated the generated knowledge graph and tag graph into Graphs, added a conventional clickable tag cloud and an initial authored story graph, and retained the complete generated graphs for inspection. Generated graph nodes now link to their notes or tags, frontmatter tag chips use the same destinations, and the static build writes one note-list page per tag under dist/tag/. Added note-link and tag-link Mermaid comments so future LLM-authored story graphs can use stable wiki names while the builder supplies their static URLs.
  • Replaced the initial linear story graph in Graphs with hardcoded but editable maps: a curated overview and separate privacy, private-AI, and business islands. These explanatory maps now coexist with the exhaustive generated graphs and can be regenerated by an LLM when important hubs change. Retyped the then-existing “Ontology” page as meta and removed the unused ontology type, so the complete knowledge and tag graphs omit that maintenance page along with the visualization page itself.

2026-07-13

  • Added a persistent title-or-updated-date sort option to the static wiki’s left note library and preserved its scroll position across full-page navigation, so recently changed notes can be found without losing one’s place.
  • Added a note-role diagram to the then-existing “Ontology” page that distinguishes knowledge-bearing, navigational, and wiki-support types. The diagram shows a typical evidence-to-synthesis flow without presenting the ontology as a mandatory pipeline, and the prose now defines index, log, and meta alongside the five substantive note roles.
  • Moved that diagram to Graphs § Ontological types, where it belongs with the other curated and generated views. The then-existing “Ontology” page retained the type definitions and now points to the shared visualization page.
  • Hardened incremental builds against incomplete dist/ trees. A current .build-stamp no longer causes a skip when a generated note, tag page, or required directory is missing. The dynamic-server smoke test now uses a temporary empty output path without moving or modifying the real dist/ tree, preventing Windows file handles from interrupting backup restoration or stranding a valid stamp.
  • Retyped Log as meta and removed the now-unused log type. The chronological maintenance history is therefore excluded from the complete knowledge and tag graphs instead of appearing as a prominent subject-matter hub.
  • Updated every curated story graph in Graphs § Story graphs to reuse the complete graph’s ontological shapes and colors. Concepts, entities, syntheses, and indexes now remain visibly distinct instead of collapsing into generic rectangles. Configured the Markdown line-length check to ignore fenced code blocks, where canonical Mermaid labels and syntax must remain contiguous.
  • Removed the initial “Knowledge base” and “Ontology” scaffold pages. Their thin definitions reflected setup prompts rather than durable subjects: the former had no grounded backlinks or sources, while the latter duplicated the canonical rules in AGENTS.md and the diagram in Graphs § Ontological types. Reduced Index to genuine entry points: Privacy venture atlas for the current subject area, plus Graphs and Log for the wiki itself. The topical pages removed from the global index remain reachable through the atlas.
  • Added Privacy gear webshop as a testable physical-product wedge in the broader privacy venture. Added Privacy gear catalog to rank useful products, require claim evidence, and exclude jammers, covert surveillance, pseudoscience, and unverifiable electronics.
  • Added Privacy gear market scan from preserved Swedish and EU competitor pages. The scan shows a fragmented market of general retailers, Faraday specialists, authentication manufacturers, and lifestyle shops, with an opening for restrained Swedish-language curation and tested kits.
  • Added Privacy-respecting ecommerce and Swedish ecommerce compliance guidance. They connect low-data shop architecture to official consumer, product-safety, privacy, tax, electronics, packaging, and jammer guidance, including the Swedish website withdrawal-function rule effective since 19 June 2026.
  • Linked the shop experiment through Privacy venture atlas, Privacy business proposals, Privacy competitors and inspirations, Privacy legal and regulatory posture, and Privacy venture roadmap so later product and service work can reuse the research.
  • Added Privacy voucher shop from preserved ProxyStore digital goods, Mullvad partner material, and official Swedish voucher, payment, crypto, AML, and digital-content sources. The design uses partner-issued fixed codes, random order secrets, optional contact, and separate physical and digital checkouts; it explicitly rejects a customer balance, crypto custody, or claims of anonymity.
  • Added Swedish privacy voucher compliance guidance to preserve the source-level distinction between a voucher sale, an electronic-money or payment instrument, and a crypto-asset service. Updated Crypto payments for privacy services, Unlinked subscription architecture, and Mullvad with the narrow merchant-acceptance and official-reseller path.
  • Added Swedish and EU privacy suppliers to rank Mullvad, Yubico, Integrity VPN, Plegium, and European physical and digital suppliers. It distinguishes company, manufacturing, distributor, and EU responsible-person location so origin marketing remains precise.
  • Expanded Privacy gear market scan to cover Swedish security, tactical, outdoor, and prepper competition. Preserved Tacticalstore, Roda Stjarnan, The General Prepper, Overlevnadsbutiken, and the Prepparen retailer directory; Ensec was reviewed but its archival redirect failed.
  • Added Defence marker sprays from Police, Kemikalieinspektionen, carrier, and Swedish supplier material. It excludes pepper and tear-gas products, requires product-specific formulation review, and adds chemical-label, aerosol-shipping, insurance, and proportional-use gates before any licence-free spray pilot.
  • Retested the previously blocked PostNord dangerous-goods source after IVPN AntiTracker was disabled. The relevant Finansinspektionen pages returned 200, while PostNord still rejected scripted requests with 403. An interactive browser passed PostNord’s Cloudflare check, confirming that the page was reachable but its bot protection still blocked the wiki snapshot importer.
  • Re-downloaded the three relevant Finansinspektionen pages after IVPN AntiTracker was disabled, preserving new immutable raw and rendered pairs for crypto-asset services, the limited-network payment-instrument boundary, and businesses covered by the AML Act. Updated Finansinspektionen crypto-asset services and Swedish privacy voucher compliance guidance to use the fresh copies.
  • Preserved PostNord’s dangerous-goods guidance as a script-free main-content snapshot captured through the interactive browser after Cloudflare verification. Defence marker sprays now links that artifact and records that it has no raw-response companion.
  • Made the first venture test explicit in Privacy venture roadmap and Privacy business proposals. The minimum viable offer is now a fixed-scope privacy and AI baseline for a small consultancy or agency: three paid pilots, no credentials or customer files, and no software, inventory, or managed infrastructure before recurring demand appears.
  • Revised that cold-start recommendation after testing the credibility problem. Added Content-led privacy shop as the lower-capital first step: Swedish buying guides, reproducible product tests, demand capture, one limited kit drop, and only then a broader shop or higher-priced advisory offer. Reframed SEK 4,900 as a later target and set a smaller paid founding-pilot test rather than assuming unknown founders can sell it cold.
  • Added GrapheneOS Pixel procurement from current GrapheneOS support guidance, Swedish public new and refurbished prices, business-retailer listings, Google Store resale terms, and Skatteverket’s VMB guidance. The comparison identifies Pixel 10a, Pixel 10, and refurbished Pixel 8 as the smallest test assortment; it also prevents margin calculations from treating VMB stock as if it carried deductible input VAT.
  • Clarified the GrapheneOS Pixel procurement price table after the mixed VAT presentation proved easy to misread. New cash price, new cost after ordinary input-VAT deduction, and refurbished cash price are now labeled separately; only Pixel 10a was genuinely cheaper new than refurbished in the observed cash-price snapshot.
  • Added Vinstmarginalbeskattning, Swedish VAT for ecommerce, and Swedish business tax map from current Skatteverket, Tullverket, and Verksamt guidance. The new path separates ordinary VAT from VMB at the inventory-unit level and keeps transaction VAT, business profit, payroll charges, and owner taxation distinct for the webshop plan.
  • Audited every maintained note for article structure, crosslinks, source use, visual opportunities, and analytical tone. The vault had no orphan notes or unresolved links; the main recurring weakness was defensive, categorical language repeated outside the notes that owned each boundary.
  • Added a neutral analytical-voice rule to AGENTS.md. Future notes should distinguish fact, law, technical invariants, business recommendations, and moral judgments; categorical language is reserved for actual prohibitions, security invariants, evidence duties, and deliberate scope decisions.
  • Added a hybrid citation policy: frontmatter sources remains the note-level provenance inventory, while descriptive Markdown footnotes or inline links map consequential claims. The renderer now styles footnotes and turns local sources/... metadata entries into clickable links. Pandoc citeproc remains deferred until the wiki has a scholarly corpus that benefits from citation keys.
  • Added Criminal-offence data under GDPR Article 10 to centralize a repeated Swedish legal and operating boundary, and Hardware security keys to give a recurring high-value product category its own deployment guidance. Monero remains an intentional wanted page because the current wiki grounds its payment role but not yet a useful protocol-level article.
  • Added comparison tables to Privacy business proposals and Private AI buyer alternatives, a data-boundary diagram to Privacy product architecture, and the current commerce branch to Graphs § Business cluster. Reframed the private-AI legal gate from moralized green-amber-red refusal language into routine, review, and separately governed specialist work.
  • Revised the agent guidance after the bounded Markdown-dialect language proved liable to suppress article depth, structural formatting, related-page creation, and wanted-page links. AGENTS.md now asks for developed, neutral synthesis in an academic-trade register; proactively considers headings, tables, Mermaid diagrams, lists, blockquotes, and callouts; welcomes grounded seed pages; and treats redlinks as a deliberate research queue rather than lint debt.
  • Re-read all 86 maintained pages under the richer synthesis, structural-formatting, graph-development, and lifecycle rules. Reclassified the vault by present usability rather than age: 63 notes are stable at their current scope, while 23 remain working because they contain active business hypotheses, unresolved legal classification, changing proceedings, supplier validation, or other material gaps. No current page is merely a seed, and no page is presently known to be stale.
  • Defined an explicit seed to working to stable workflow in AGENTS.md and retained stale as an override for claims known unsafe to rely on. Added optional review_after scheduling to 51 time-sensitive notes, visible review metadata in the renderer, just status for the maturity queue, and just review for stale and due-review discovery.
  • Revisited the earlier content audit under the more liberal linking rule. Added a ten-case comparison table and incident redlinks to Case for privacy and security, a route-and-evidence table to Private AI competitors, and a lead-comparison table to Swedish and EU privacy suppliers. Added wanted-page links for reusable incidents, companies, products, and technical projects throughout the case, competitor, market, and supplier notes; the wanted queue now contains 51 prospective articles.
  • Added missing source provenance to Ethics of resisting lawful extraction, linked existing private-AI entity pages where plain names had hidden them, and updated Private AI venture validation to use the current routine, review, and specialist legal-role language.
  • Clarified alias semantics in AGENTS.md: aliases support discovery and natural display text, but maintained wikilinks now use canonical targets for agreement between Markdown Oxide and the static renderer. Normalized the five historical alias-target links in this log and taught the linter and just fix to prevent their return.
  • Reclassified discreet cameras and audio recorders in Privacy gear catalog from a moralized “do not sell” category to a dual-use specialist category with explicit decision gates. Added Covert recording and evidence in Sweden to distinguish participant recording, recording conversations between others, intrusive photography, camera-surveillance rules, and evidentiary use. Tightened AGENTS.md so an “explicit scope decision” cannot hide an unexamined moral judgment, and neutralized similar categorical framing in Privacy legal and regulatory posture, EU private communications law, Case for privacy and security, and Privacy competitors and inspirations.
  • Audited the maintained vault for callouts and task lists. Before this pass, no note used tasks and only the newly added Covert recording and evidence in Sweden used a callout. Added semantic callouts to legal, technical, procurement, and product-gate notes; exposed executable validation, launch, and supplier work as task lists in Private AI venture validation, Privacy venture roadmap, Privacy gear webshop, Swedish and EU privacy suppliers, and the covert-recording synthesis. Updated AGENTS.md so agents actively consider these structures rather than treating renderer support as sufficient.
  • Corrected the new covert-recording synthesis to define IMY and NJA through the local abbr frontmatter feature. Rewrote its redlinks as durable subject prose rather than describing unresolved targets as “future pages” or “wanted pages.” Clarified the same rule in AGENTS.md: a subject-note sentence containing a redlink should not need revision when the target article is created.
  • Added local abbreviation metadata to 78 maintained notes so the static renderer can emit semantic HTML abbr elements with all-small-caps styling while the source remains ordinary Markdown. Documented quoted YAML abbr entries as the house style in AGENTS.md, retained Markdown Extra definitions for compatibility, and added lint checks for malformed or conflicting definitions.
  • Added the mixed-case abbreviation MiCA to the four maintained notes that use it after the initial all-uppercase corpus audit missed it.
  • Audited exact abbreviation surface forms after CASPs and DAC8 exposed the same gap for plurals and mixed alphanumeric terms. Added local plural definitions where maintained prose uses them and clarified the exact-match rule in AGENTS.md.
  • Resolved the complete 55-page wanted queue with source-grounded entity and concept notes. Added incident and enforcement pages for 23andMe data breach, Afghan relocation data breach, Ashley Madison data breach, Equifax data breach, Grindr data-sharing enforcement, OPM data breach, Pegasus spyware, SpyFone, Vastaamo data breach, and VTech data breach from regulator, government, court, and investigative sources. Grounded the Swedish evidence-law branch through Olovlig avlyssning, Kränkande fotografering, Free evaluation of evidence in Sweden, and Evidence preservation for abuse and workplace disputes using the current Penal Code, Supreme Court material, and police guidance. Added current product and supplier cards for addy.io, Apple Advanced Data Protection, BeProtected, Clas Ohlson, Cryptomator, DataProtect Sverige, DeleteMe, Disklabs, Ensec, Faradaybags.cz, Google client-side encryption, Incogni, Integrity VPN, IVPN, Kjell & Company, Koofr, Little Snitch, Loomis, Monero, Nitrokey, Obscura, OffGrid, Optery, Plegium, Roda Stjarnan, Tacticalstore, Threema, Tresorit, Tyloona, Urban Privacy, and Yubico. Completed the private-AI and Nordic enterprise-security branches with Confident AI, Enkrypt AI, Limina, llama.cpp, MLX, Ollama, Open WebUI, mnemonic, Truesec, and WithSecure. Current product pages now carry scheduled reviews; immutable incident pages distinguish observed findings from venture lessons; and removal, Faraday, local-AI, and managed-security pages state the boundaries that would make their strongest marketing interpretation false.
  • Corrected the abbreviation metadata across the 55-page wanted expansion. Added missing local definitions for legal, geographic, technical, and institutional terms; replaced singular definitions where the prose uses only plural forms; removed definitions that the note does not use; and normalized the FIDO expansion to the established house style.
  • Defined currency codes locally in 23andMe data breach, Equifax data breach, Grindr data-sharing enforcement, VTech data breach, and this log. Clarified AGENTS.md so maintained prose consistently treats ISO currency codes as abbreviations instead of relying on a discretionary exception.
  • Imported and read Thérèse Juel’s Fällda för sexövergrepp and Dömda, together with the official Felaktigt dömda and Rättssäkerheten i brottmål reports. Preserved research on confirmation bias, political lay-judge effects, offence-specific evidence patterns, the current Code of Judicial Procedure, remand isolation, police-arrest overflow, double occupancy, the 2026 proposal to legalize capacity workarounds, and current oversight of lay judges. Built Rättssäkerhet in Swedish criminal cases around uncalibrated proof, false-accusation uncertainty, incomplete investigation, political adjudication, coercive remand, and mobile-forensics selection risk. Added Duress credentials and coercive extraction to distinguish a user’s moral decision under coercion from a privacy provider’s operational boundary, and connected the result to Case for privacy and security, GrapheneOS, and Privacy product architecture. Preserved Lambertz’s primary SvJT article as Kvalitetssäkring av bevisprövningen i brottmål and developed Probabilistic interpretations of beyond reasonable doubt to separate the popular 98-percent gloss, its one-in-fifty implication under calibration, and the absence of calibration in actual adjudication. Added the ASA statement on p-values and used the criticized p < 0.05 convention to distinguish a defined but frequently misinterpreted statistical rule from the legally consequential 98-percent figure that has no reproducible courtroom operation.

2026-07-14

  • Generated ten mockups for the wiki application: an editorial workbench, a more speculative knowledge atlas, a rendering-component board, two dark-theme studies, two Mermaid studies, and three deliberately ungrounded interface proposals. The ungrounded pass used no screenshots or existing CSS as visual input; it explored a spatial reading desk, an editorial folio, and a graph-native focus environment. Used the editorial direction to redesign the live renderer with a calmer, wider reading surface, a recoverable desktop library-collapse control, flatter navigation and article surfaces, a restrained title scale, compact metadata, clearer outline state, type-aware connection cards, semantic callout colors, and copy controls for code blocks. Added locally persisted note pinning, type-specific icons in navigation, metadata, pins, and backlinks, a blue-black archival dark palette, and a more deliberate Mermaid theme with curved routes, clearer clusters, edge labels, and an ontology legend for large graphs. These changes selectively combine useful ideas from several mockups rather than reproducing any one generated composition. Preserved the real Index, Graphs, and Log landmarks instead of implementing speculative destinations invented by the more radical mockup.
  • Corrected Rättssäkerhet in Swedish criminal cases so it no longer treats an independent witness as ground truth. Added Witness reports are not ground truth to distinguish independence of people from independence of errors and connect the JK report’s memory-psychology analysis with research on co-witness contamination and procedure-dependent identification accuracy.
  • Changed the Tags § Tag cloud from a linear frequency scale to a bounded logarithmic scale. Low- and middle-frequency tags now differ visibly in size, while the largest tag retains the previous maximum size.
  • Changed the static wiki’s “Updated” sidebar sort to use each source note’s actual filesystem modification time instead of its frontmatter updated date. The displayed note metadata still reports the maintained frontmatter date.
  • Repaired the Graphs § Ontological types Mermaid source by separating shaped-node declarations from graph edges. The renderer now isolates Mermaid failures per diagram, and lint rejects the invalid inline declaration form, so one malformed graph cannot suppress every graph on a page. Raised Mermaid’s explicit edge limit to accommodate the complete knowledge graph’s 645 resolved links.
  • Added Boomers and democracy as a source-grounded Swedish assessment of moving cohort capture, prompted by the imported ChatGPT conversation of the same name. Added Moving cohort capture, Age-period-cohort identification, and Welfare-state dualization to separate the proposed mechanism from its methodological limit and the better-established insider–outsider pattern. Preserved SCB demographic and election material, Swedish pension guidance, and scholarship on ageing democracy and Social Democratic retrenchment; created six source cards so later work can audit the demographic premise, the 2006 electoral evidence, pension transition rules, and the main alternative explanations.

2026-07-15

2026-07-16

  • Expanded Index with distinct entry points to Rättssäkerhet in Swedish criminal cases and EU age verification and Internet privacy. The first opens the Swedish criminal-procedure and evidence branch; the second opens the age-assurance, Internet-governance, and pornography-literacy branch without duplicating their onward reading paths.
  • Added a Cloudflare Pages deployment adapter for Publishing the wiki privately. It preserves Index as the note’s canonical page, redirects the host root to /Index, and adds noindex and defensive response headers at deployment time instead of changing the wiki renderer around one host’s root convention. The recipe prints the stable production alias after Wrangler’s immutable per-deployment URL so the shareable address is unambiguous.
  • Added Publishing the wiki privately to distinguish obscure-link noindex publication from actual identity-gated access. The recommendation uses Cloudflare Pages for the static build, Cloudflare Access for sensitive or whole-vault sharing, and treats ChatGPT Sites as a future workspace-bound option rather than the current free default in the European Economic Area. Preserved the decisive Cloudflare and Google documentation in Cloudflare Pages and Access official publishing documentation; the OpenAI help page returned HTTP 403 to the repository downloader, so ChatGPT Sites official publishing documentation records the official live source and that acquisition limitation.
  • Added Redact.dev as a social-account bulk-deletion competitor distinct from broker-removal subscriptions. Preserved its current architecture, privacy, terms, and pricing pages in Redact.dev official product and policy pages. The assessment records the attractive local-processing design, the vendor’s explicit best-effort deletion limits, and an unresolved June-July 2026 telemetry finding involving a resolvable X account identifier.
  • Clarified the EU-law boundary in Redact.dev: a compliant platform should make third-party automation unnecessary for permanent account erasure, while selective and scheduled cleanup remains a convenience case. Preserved EDPB deceptive design patterns in social media guidelines as the governing interface and account-closure guidance.
  • Added Nym VPN and Portmaster and SPN as distinct network-privacy competitors. The comparison separates Nym’s delayed, cover-traffic mixnet from SPN’s application-aware onion routing and records that IVPN now develops and operates Portmaster and SPN.
  • Preserved current official product, pricing, architecture, and ownership pages in Nym VPN official product pages and Portmaster SPN official product pages.
  • Preserved Analysis and attacks on the reputation system of Nym and When mixnets fail so the Nym analysis includes current independent counterevidence on reputation attacks, centralization, compromised relays, and path-selection tradeoffs.
  • Updated Privacy competitors and inspirations, Split-trust network privacy service, and IVPN to connect the new evidence to the venture’s product landscape.
  • Broadened Alternatives to age verification from a pornography-centred comparison into a five-category analysis of content, conduct, contact, consumer, and cross-cutting risks. The revision separates universal safer design, feature- or transaction-level proof, and service-wide exclusion.
  • Added Less harmful social media to develop alternatives to blocking for compulsive design, sleep disruption, social comparison, harmful recommendation, harassment, grooming, commercial manipulation, spending, privacy, crisis content, and artificial-intelligence companions. It also records the social and developmental functions that a proportionality analysis must count.
  • Preserved the Commission’s 2025 protection-of-minors guidelines and the 2026 Joint Research Centre systematic review in DSA protection of minors guidelines and JRC social media and youth mental health review. Added Ofcom protective social media defaults trial with an explicit boundary between persistent settings and demonstrated real-world harm reduction; the official report rejected automated archival requests with HTTP 403.
  • Updated Index so its existing age-verification entry describes the branch’s broader work on safer online services rather than presenting pornography literacy as its defining destination. The branch remains one top-level entry because EU age verification and Internet privacy already routes onward to Alternatives to age verification and Less harmful social media.
  • Audited inline code spans across all maintained notes. Replaced code styling used as quotation marks, term styling, and mathematical notation with straight quotes or italics, while preserving code spans for machine-readable literals. Corrected the stale spaced-tag example in Tags and clarified the inline-markup policy in AGENTS.md.

2026-07-17

2026-07-19

2026-07-20

2026-07-21

  • Merged the adolescents tag into adolescence, the sole near-duplicate pair the tag report flagged. The two names split one facet — the developmental stage — across the population plural and the stage noun. adolescence is the established name (11 notes against 3, and consistent with the vault’s other stage nouns children, generations, and ageing), so the three social-media notes JRC social media and youth mental health review, Less harmful social media, and Which social-media designs reduce harm for young users now carry adolescence. Ordinary prose uses of the word “adolescents” were left untouched.
  • Merged the evidence-review tag into evidence-synthesis. Both name the same methodological facet — a note that is or reports a systematic research review — and evidence-synthesis already carried eight such notes against evidence-review’s one, JRC social media and youth mental health review, which is itself a systematic review and now joins the others. The sibling tags evidence-law (a legal field) and evidence (evidentiary material and proof) were left in place; they share a leading word with evidence-synthesis and nothing else.
  • Merged the cybersecurity tag into security. The two named one idea applied to disjoint notes: cybersecurity sat on three incident-response firms mnemonic, Truesec, and WithSecure plus X-rated compliance theater, while security carried the conceptual and product notes. Because they shared no note, neither the spelling check nor the co-occurrence view could surface them; the pair was found by reading the counts. The firm-specific facet stays on the incident-response tag. The ordinary English word “cybersecurity” in prose was left untouched.
  • Merged the crypto tag into cryptocurrency. This is the spelling-divergent synonym applied to disjoint notes that the co-occurrence and near-duplicate views cannot surface: crypto sat on five payment and compliance notes Crypto payments for privacy services, Finansinspektionen crypto-asset services, Privacy voucher shop, ProxyStore digital goods, and Swedish privacy voucher compliance guidance, while cryptocurrency carried the coin and protocol notes Bitcoin, Monero, Lightning Network, and their comparisons. A reader searching either wants the other set, and the payment facet those notes share is already carried by the payments tag, so crypto added no navigation of its own. The distinct payment facet stays on payments.
  • Triaged two single-use tags the earlier passes had left in the tail. Retired source, which sat on OfficeChai report on Taalas ChatJimmy and merely restated that note’s source type, separating nothing. Expanded reseller from one note to two: ProxyStore digital goods describes itself as a German digital-goods reseller and is the genuine parallel to Integrity VPN, so it now carries the tag. Reading the candidates disciplined the expansion — most notes mentioning reselling discuss it as a plan or list their own resellers rather than being reseller entities, so the tag reached exactly one further note, not the dozen it first appeared to fit.
  • Merged the legal-reasoning tag into legal-theory. Both sat on the single note Scandinavian legal realism, a school of jurisprudence, where neither the how-judges-reason sense nor the what-law-is sense was being drawn as a distinction — the two names redundantly labelled one note. This is a spelling-divergent synonym the lexical near-duplicate check cannot catch; it was found by reading the counts. legal-theory is the more accurate name for the subject and stays as an early tag the vault may grow into.
  • Added privacy to Google client-side encryption. The near-miss containment view named the note twice — the sole encryption note and the sole cloud-storage note lacking privacy — and the repetition marked it as a lost tag, not a genuine one-off. A customer-controlled encryption layer that keeps content confidential from the cloud provider is a privacy mechanism, so the note now joins every other encryption and cloud-storage note under privacy.
  • Merged the data-protection tag into data-governance. data-protection sat on the single note Connected-vehicle data clearing, a service that erases personal data from vehicles before resale, which is the same GDPR personal-data field as the four data-governance notes Criminal-offence data under GDPR Article 10, Data portability, High-consequence data, and Privacy-respecting ecommerce. The two names are a spelling-divergent synonym applied to disjoint notes, invisible to the co-occurrence and near-duplicate views; data-governance is the established, better-populated name. The ordinary hyphenated phrase “data-protection” in prose and URLs was left untouched.
  • Merged the decentralized-communication tag into decentralization. The longer name was a two-note facet-split of the five-note tag, carried only by Secure Scuttlebutt and Secure Scuttlebutt protocol sources, which already hold peer-to-peer and distributed-systems. Its siblings AT Protocol, Nostr, and Session use the shorter decentralization, which now carries all seven notes and separates nothing the longer name did.
  • Expanded recording from one note to two. The used-once view ranked it high on title mentions, and reading confirmed the omission: Covert recording and evidence in Sweden is titled for its subject and is the evidentiary synthesis paired with Olovlig avlyssning, the sole recording note, yet lacked the tag. The two now navigate together. Examined but deliberately left this pass: the cryptocurrency + payments alias? pair, where payments is an under-populated broad tag a future card or bank-transfer note would carry without cryptocurrency, so the two are not synonyms and neither should absorb the other; the business + privacy and competitor + privacy near-miss rows, whose exceptions ENISA threat landscape 2025, Swedish business tax map, and Mistral AI are correct one-offs outside the narrower tag, not lost tags; and the assurance and Google singletons, whose title mentions came from the distinct age-assurance tag and from Google-as-subject rather than Google-as-vendor notes.
  • Expanded metadata from one note to four. It was the highest title-signal singleton left in the used-once list (two titles, forty-five body mentions), and reading confirmed the omission: Metadata-minimizing communication product gap and Metadata-minimizing messengers source bundle both name the subject in their titles yet lacked the tag, and Secure and decentralized communication stacks frames its analysis around which servers retain metadata. The same discipline narrowed the reach: Signal and Nym VPN discuss metadata as a dimension rather than as their subject, so they were left untagged.
  • Ran a further tag pass and resolved the security cluster the earlier passes had not reached. Merged information-security into security: a spelling-divergent synonym on the disjoint pair DataProtect Sverige and Loomis, Swedish security-service entities a security search should return. Merged security-analysis into security-audit: two singletons naming one idea, the formal review of a messaging protocol’s security, on Jaeger and Kumar group chat encryption analysis and White Noise and Marmot security reviews, which already shared messaging and cryptography. Merged legal into law: a one-note synonym of the established 37-note tag, on Post-seizure digital recovery.
  • Retired rights as ornamental. It sat on Justifications for caring about privacy rights, duplicating a word in that note’s own title and naming a subject too broad to navigate; the specific facets human-rights and labour-rights carry the real distinctions.
  • Reopened the competitor + privacy near-miss the earlier pass had closed and reached the opposite verdict on Mistral AI. The earlier entry grouped it with ENISA threat landscape 2025 and Swedish business tax map as correct one-offs, but those two are genuinely non-privacy notes, while Mistral is the same class as every other AI competitor tracked for the private-AI venture — Anthropic, OpenAI, Google Gemini, and Microsoft Copilot all carry privacy. The omission reads as a lost tag, not an exception, so Mistral now carries privacy. The user confirmed reversing the earlier session’s judgment.
  • Added privacy to Dumpen court monitoring weeks 21 and 22 2026. Its direct sibling of the same genre, Dumpen selected judgments, carries privacy; the monitoring batch had lost it. The privacy + public-records near-miss named it as the one exception.
  • Examined and left this pass: the network-privacy / network-security pair, where the note sets show a real distinction — VPN and split-trust products against mixnet attack analysis — not one idea spelled twice; the evidence, evidence-law, and evidence-synthesis trio, three genuine facets sharing only a leading word; and the remaining title-signal singletons Proton, Google, mobile, evaluation, detention, and prisons, where the vendor-tag convention is too inconsistent to call an omission or the tag names a distinction the broader tag does not draw.
  • Ran a whole-vault structure audit from just links and just clusters. The topology was already legible; the work was reconnecting notes the reports flagged as stranded. Every genuinely unreferenced note gained an inbound link from the subject note that should have cited it: the three product-gap source bundles (Metadata-minimizing messengers source bundle, Private household infrastructure source bundle, Private knowledge tools source bundle) are now cited by the synthesis notes they feed; IMY legal position on people-search services anchors the regulatory timeline in Utgivningsbevis and people-search services; European Digital Identity Wallet overview is cited from Privacy-preserving identity integration; and OfficeChai report on Taalas ChatJimmy now appears in Local AI as the model-specific-silicon contrast it was written to draw.
  • Reconnected Redact.dev, a developed entity note nothing linked to. It occupies the in-account-deletion niche beside the broker-removal services DeleteMe, Incogni, and Optery, which named each other but not it; all three now point to it, and Privacy competitors and inspirations adds it to the data-removal market.
  • Gave the Connected-vehicle data clearing business proposal an outbound link to Privacy business proposals, clearing a dead end whose evidence discipline is shared across the portfolio.
  • Checked but did not merge the three Private AI hubs (Private AI strategy, Private AI competitors, Private AI buyer alternatives) and the Dumpen / data-removal-rights boundary: each pair answers a distinct question for a distinct reader and is already well bridged, so crosslinking, not merging, is correct.
  • The two “orphan” publishing source cards (ChatGPT Sites official publishing documentation, Cloudflare Pages and Access official publishing documentation) are correctly cited by the meta note Publishing the wiki privately; the graph excludes meta links, so their orphan status is a report artifact, not a defect.
  • Imported the DoldAdress.se homepage (DoldAdress homepage, artifact sources/2026-07-21-doldadress-hemsida.html) and wrote the entity note DoldAdress, a Swedish subscription service that suppresses personal data across the domestic people-search sites and de-indexes Google results.
  • Placed it against the vault’s existing legal foundation: its removal targets are the certificate-bearing directories analysed in Utgivningsbevis and people-search services, so its “Länkgaranti” runs into the outcome-guarantee problem that EU and Swedish data removal rights already flags.
  • Added it to the data-removal section of Privacy competitors and inspirations as the Swedish-local incumbent, whose missing removal-evidence report marks the entrant’s opening.
  • Imported the air.ms homepage and privacy page (artifacts sources/2026-07-21-air-ms.html and sources/2026-07-21-air-ms-privacy.html) and wrote the source note Air secure messenger, a free EU-hosted MLS messenger from the MLS and OpenMLS authors Phoenix R&D that registers without a phone or email and encrypts push tokens.
  • Placed Air in Secure and decentralized communication stacks with a matrix row and a prose subsection beside the metadata-minimizing messengers, and backlinked it from Messaging Layer Security as a current MLS deployment alongside White Noise. The recurring caveat applies: strong protocol provenance, but an early and not-yet-independently-mature deployed system.
  • Left Phoenix R&D and OpenMLS as redlinks; the only evidence this session is Air’s own marketing page, so they enter the wanted queue rather than becoming vendor-sourced notes.
  • Evaluated the PrivSec.dev “F-Droid Security Issues” article for fairness. Imported it (artifact sources/2026-07-21-privsec-fdroid-security-issues.html) and wrote the source note PrivSec.dev on F-Droid security issues, noting that its July 2026 timestamp hides mid-2022 content.
  • Wrote the concept Source-to-binary correspondence to name the reusable idea the exchange turned on: signing proves publisher continuity, not that a binary matches inspectable source, and only Reproducible builds bridge the gap. This is the property the trusted-party count silently prices at zero.
  • Wrote the synthesis Android app distribution trust models, comparing Play Store, developer-signed, and F-Droid channels on the correspondence axis and testing the PrivSec critique against it: durable complaints (update latency, low client target API, stale dependencies) survive, but the “prefer Play Store” conclusion does not, since it assigns no value to source-to-binary verification. Recorded that ~21% of F-Droid main-repo apps were reproducible and developer-signed by end of 2025, with a status dashboard since May 2025.
  • Left Reproducible builds and Android Package as redlinks; both are nameable and reusable but not grounded enough this session to write vendor-neutral notes.
  • Researched two follow-up claims and corrected the notes. The “outdated infrastructure” argument rests only on a 2022 window when F-Droid’s build VM ran end-of-life Debian Stretch; F-Droid replaced its decade-old core server in 2025 and moved build orchestration to Buildbot, so the claim now cites a resolved incident as current state. Recorded this in PrivSec.dev on F-Droid security issues.
  • Added a signing-key custody section to Android app distribution trust models: Android enforces signature continuity, so key custody is the top target. F-Droid uses one key per app by default, not one global key, so a leaked key compromises one app; but all signing runs on a single air-gapped machine, so compromising the host exposes the whole keystore. Many keys, one vault. Reproducible builds are the backstop F-Droid’s own security model names against a key-holding attacker.
  • Resolved the Phoenix R&D and OpenMLS redlinks left by the Air import, this time grounding them in non-vendor primary sources. Confirmed against RFC 9420 that Phoenix R&D’s Raphael Robert is a listed MLS author, and against the Phoenix R&D and Cryspen blogs that OpenMLS is co-maintained by the two companies, correcting the Air note’s earlier “maintains” to “co-maintains” in three notes.
  • Wrote OpenMLS (entity): MIT-licensed Rust RFC 9420 implementation, first independent audit by SRLabs sponsored by the Sovereign Tech Agency, report v1.1 dated 2026-03-03, eight issues (one High), seven fixed in crate 8.1/7.3. Preserved the audit summary as sources/2026-07-21-openmls-security-audit.html.
  • Wrote Phoenix R&D (entity): German public-interest GmbH sitting at three layers of the same technology — MLS spec authorship, OpenMLS co-maintenance, and the Air product — funded by public-interest grants (OTF Project Phoenix, NLnet, Prototype Fund, Sovereign Tech Fund) rather than advertising. Preserved the company page as sources/2026-07-21-phoenix-rd.html.
  • Backlinked both from Messaging Layer Security as MLS implementations and deployments. Kept the grant funders as plain prose, not redlinks: their only vault relevance so far is these two entities.
  • Corrected a protocol-versus-product conflation flagged by the user: “the Signal Protocol” is a reusable cryptographic construction (PQXDH plus Double Ratchet, reference library libsignal), distinct from Signal the app, and is deployed by WhatsApp, Google Messages RCS, Facebook Messenger, and Skype. Added a “protocol is not the product” section to Signal and a “Comparison with the Signal Protocol” section to Messaging Layer Security, framing MLS against the Signal Protocol as building block against building block. The distinguishing axis is interoperability standardization, not openness: libsignal is open source but not a wire-interop standard, whereas MLS (RFC 9420) is a multi-vendor IETF standard. Grounded the RCS example in the GSMA Universal Profile 3.0 announcement (MLS chosen for interoperable iPhone-Android E2EE, March 2025, shipped 2026), and the non-Signal deployments in Signal’s own WhatsApp-completion post and the Signal Protocol overview. Left the popular “proprietary/single-vendor” framing out as imprecise.
  • Sharpened the interop framing in the same two notes after the user caught an internal tension: the first pass said Signal Protocol “adopters run the same library,” which reads as “single implementation” and contradicted the adjacent mention of reimplementations. Corrected both to state that independent Signal Protocol reimplementations do exist (the design is not copyrightable), and that what it lacks is a specified wire-format interoperability standard, not implementation diversity. Reframed the MLS contrast as specification conformance rather than a shared codebase being the basis of interop.
  • Added a “Cryptographic guarantees” subsection to Messaging Layer Security. Both protocols provide confidentiality, forward secrecy, and PCS asynchronously; the note now states the two distinctions that were previously only implicit across the vault: the Double Ratchet heals per message while MLS heals per epoch (so PCS is only as frequent as group updates), and the comparison flips with conversation size — Signal Protocol is the stronger 1:1 guarantee (finer granularity plus deployed post-quantum ratchet), MLS the cleaner large-group guarantee (native tree-based FS/PCS versus an application-level sender-key construction where group PCS is hard). Cited the Double Ratchet spec and RFC 9420 for the granularity claim, and kept a callout tying it back to Jaeger and Kumar group chat encryption analysis: protocol-level FS/PCS does not certify a deployed group.
  • Researched the user’s open-standard federated e2ee email proposal and wrote Open standard for federated e2ee email. Grounded the PGP metadata critique in Latacora’s analysis, traced why Autocrypt and WKD did not close the gap, and used Delta Chat’s Chatmail relay as the closest working reference implementation of the proposed s2s/c2s split. Tested the proposal against Moxie Marlinspike’s federation-means-stasis argument and found Chatmail a partial counterexample, but only by accepting a narrower, opinionated profile than the user’s IMAP-proxy compatibility goal implies; flagged that full IMAP compatibility and end-to-end encryption to a legacy client are close to mutually exclusive. Linked the OpenPGP Email Summit as the standing venue where vendors including Proton already work this problem. Added a “Federated encrypted email” subsection to Privacy market gaps.
  • Researched the user’s EU-finetuned Chinese open-weight model proposal and wrote EU-finetuned Chinese open-weight models. Confirmed the export-control premise is live and accelerating, with Moonshot AI’s Kimi K3 (mid-July 2026) the direct trigger for the current US debate over restricting Chinese models at the software layer. Found the censorship-removability question only partly settled: Khoury College research shows DeepSeek-R1’s refusals bypass at inference time without retraining, suggesting an RLHF-stage mechanism, while a separate finding that the behavior “persists across languages and transfers to distilled models” cautions that it may survive light finetuning: framed these as two distinguishable claims (refusal removal versus leaning removal) needing separate evaluation. The most consequential finding was competitive, not technical: Venice already proves the “privacy-positioned, uncensored open-weight model” business model works at $1B valuation and $70M ARR, and a second startup, eustella, is already executing the user’s specific EU-hosted, Chinese-model, sovereignty-positioned strategy, which repositions the opportunity from greenfield to a named-competitor problem. Added a “Chinese open-weight models, finetuned for Europe” subsection to Private AI competitors.
  • Verified the TEE question against the existing Confidential AI computing and Private AI trust boundaries notes, which already correctly state that TEEs provide attested confidential processing, not zero-knowledge or E2EE. Added the October 2025 TEE.fail attack (Intel TDX, AMD SEV-SNP, and NVIDIA GPU confidential computing, physical DDR5 interposition, under $1,000 in equipment, attestation keys extracted) to both notes’ residual-risk sections as concrete, dated evidence for the previously abstract “side channels are real” claim. Added a direct answer in Private AI trust boundaries to what a TEE changes versus a privacy policy: the provider still reads plaintext to compute either way, and what changes is that the customer’s client verifies fresh evidence of what code is running before it will send anything, making substitution by unauthorized administrators or components detectable rather than merely promised against.
  • Added Obsidian-style math (inline and
    block
    ) to the wiki’s supported syntax. The renderer converts LaTeX to MathML at build time with the pure-Python latex2mathml library rather than vendoring MathJax or KaTeX, matching the constraint that already drove the Mermaid vendoring choice: notes must render correctly opened directly from dist/ over file://, with no script and no network access. Math extraction runs before Markdown/smarty processing so LaTeX characters like _, ^, and -- are not reinterpreted as Markdown, and the inline-math regex requires no whitespace immediately inside the delimiters — the same rule Obsidian itself uses — so that currency prose such as “$1B valuation and $70M ARR” is not misread as one math span. Rewrote the two pseudo-code formulas in GrapheneOS Pixel procurement (the VAT net-cost calculation and the device margin breakdown) as real $$ block math. Documented the new syntax and its build-time-MathML behavior in AGENTS.md.
  • Fixed two math-rendering defects found by reviewing the rendered GrapheneOS Pixel procurement page. .math-block set only overflow-x: auto, and CSS computes the other axis to auto as well whenever it isn’t visible, so any block whose MathML was a few pixels taller than its line box (any fraction) got a spurious vertical scrollbar; added an explicit overflow-y: hidden and a little vertical padding. Separately, the margin-rule formula’s seven \text{}-labelled subtracted terms made a single equation too wide for the note column with no way to break it, since latex2mathml does not lay out multi-row aligned/array environments; replaced it with a short equation, device contribution = customer price excluding VAT - sum of costs, followed by an ordinary bullet list of the cost components, which reads better than a wide subtraction chain regardless of column width.
  • Made block math visually distinct from surrounding prose instead of sitting in it unstyled. First pass stacked a border, an accent text color, and a font swap to sans-serif on top of a tinted background, which read as too heavy for a two-line formula; kept only the tinted panel and dropped the rest, so it stands apart the same way a code block does without shouting. Inline math is unstyled and reads as part of the sentence.
  • Fixed the tinted panel’s content looking off-center vertically. The box’s own padding was already symmetric, but centering it with text-align on a block-display MathML element left the fraction’s ink sitting closer to one edge than the other; switched the panel to a flex container with align-items: center so it centers the math element’s actual rendered box instead of its line box.
  • Checked a report that the panel’s left/right padding looked uneven: measured it directly and the padding is equal on both sides in both formulas. What reads as off-center is the equals sign, since an asymmetric expression’s two halves are different widths and centering the whole span puts the visual focal point left of the box’s middle — not a padding bug, and left as is. Switched the math font to the body serif instead of the browser’s default math font so it reads as part of the note’s typography.
  • Replaced the body serif as the math font with Libertinus Math. Source Serif 4 has no OpenType MATH table, so the fraction rule and the summation sign in GrapheneOS Pixel procurement were visibly mis-sized and mis-positioned — browsers need a real math companion face to lay out fractions, stacked symbols, and stretchy delimiters correctly. Libertinus is designed to pair with Linux Libertine/Biolinum, whose proportions are close to Source Serif’s, and is SIL OFL licensed like the wiki’s other bundled fonts. Vendored LibertinusMath-Regular.woff2 under share/fonts/; the Libertinus project’s own prebuilt WOFF2 release asset is missing the MATH table, so this copy was recompressed from the release’s OTF with fontTools, which preserves it, documented in share/fonts/README.md. Added a --serif-math CSS variable (Libertinus Math, falling back to other MATH-table fonts a reader might already have installed) rather than overloading --serif, since Libertinus Math’s wide-set glyphs are unsuitable for body text.
  • Ran a whole-vault structure audit (just links and just clusters across all 22 clusters and 325 subject notes). The vault turned out to be unusually well-maintained: no cluster needed a merge, and the two just links orphans, ChatGPT Sites official publishing documentation and Cloudflare Pages and Access official publishing documentation, are false positives caused by their only inbound link living in the meta-typed Publishing the wiki privately, which the orphan/dead-end reports deliberately exclude. Findings that were real:
  • Added five missing crosslinks from dead-end source cards to the entities they describe by name: IVPN Unlinked Access to IVPN, Obscura trust 2-party relays and QUIC to Obscura, and Analysis and attacks on the reputation system of Nym, When mixnets fail, and Nym VPN official product pages to Nym VPN. This shrank the wiki-wide dead-end count from 9 to 4.
  • Trimmed two passages that restated another note’s canonical content instead of linking it: the proportionality test in Case for privacy and security (canonical version stays in Ethics of resisting lawful extraction) and the Article 10 operating recipe in Sweden as a privacy venture base (canonical version stays in Criminal-offence data under GDPR Article 10).
  • Resolved an age-tier inconsistency in the pornography and age-verification cluster: three notes each stated their own age-graduation model with different granularity. Less harmful social media’s four-tier model (prepubertal children, early adolescents, older adolescents, young adults) is now the canonical version; Alternatives to age verification and Pornography literacy and harm reduction link to it instead of restating a coarser two- or three-tier version of their own.
  • Fixed an uncited claim in Pornography literacy and harm reduction about authentic-looking sex not proving ethical production; it now points to the sourced version in Better pornography production.
  • Added a redlink pair, Ethereum, Solana, and Zcash, in Cryptocurrency design comparison, which discussed all three systems at length without linking any of them.
  • Added roughly a dozen smaller missing crosslinks across the privacy, private-AI, and Swedish-legal clusters, including Data portability (previously linked from nowhere despite four notes describing exactly its subject), Private AI routing and Private AI trust boundaries (the same trust ladder stated as a taxonomy and as a decision procedure, previously not cross-referencing each other), and Swedish remand detention and restrictions to Ethics of resisting lawful extraction.
  • No merges were made anywhere in the vault. Several suspected duplicate pairs — the three “Private AI” competitor/service/buyer notes, the three confidential-computing/trust-boundary/zero-knowledge notes, and the near-identically named Rättssäkerhet in Swedish criminal cases / Rättssäkerheten i brottmål / Dömda — were read in full and confirmed to be correctly distinct, cross-linked notes rather than restatements.
  • Ran another tag pass and re-verified the earlier pass’s report reached the same resolved state, with one correction: the user confirmed that Google should tag any note genuinely about Google as an actor, overturning the prior pass’s declined expansion. Added Google to Google Gemini (Google’s own AI product) and Google private-information removal policy (a source card documenting Google’s specific removal policy), joining the already-tagged Google client-side encryption. Left it off TU and RE v Google C-460-20, GC and Others C-136-17, and Conseil d’État 401258: those three are about the CJEU’s search-engine-de-referencing doctrine, which generalizes to any search engine and names Google only as the party in one dispute, so tagging them would misstate the note’s subject rather than reflect an omission.
  • Recorded naming candidates for the wiki and repository in the new Naming the wiki meta note. A naming session floated five threads — architecture, contents, Obsidian riffs, sovereign riffs, and elements named after Ytterby gruva — with Scandium, Allodial, Loomstead, and Provenance as frontrunners. No name is chosen yet; the note carries tasks to pick a direction and check the choice for software collisions.
  • Added a second naming session’s candidates to Naming the wiki: Redlink, Wayfinder, and Seedbed under architecture-grounded names; Rättssäkerhet, Skyddade, and Confidant under content-grounded names; Flint under Obsidian riffs; and Sovran and Suvern under sovereign riffs. Added Sovran as a frontrunner for a dictionary-attested sovereign-thread name with no diacritic and no risk of reading as a typo.
  • Added Unmasking pseudonymous public figures after research into the 2026 Reuters investigation naming Banksy as Robin Gunningham and the ethical debate it produced. The note applies the wiki’s existing purpose-versus-form distinction for public-interest identification, developed for Swedish court reporting in Legal Newsdesk Sweden C-199-24 and Naming convicted people in Swedish media, to a non-Swedish case: a publisher’s own judgment that unmasking serves the public is a separate question from whether the identification is accurate. Recorded that Banksy’s UK legal exposure — graffiti is criminal damage, with a Grade I listed building able to lengthen sentencing — gives criminal prosecution an independent, non-journalistic path to the same disclosure outcome, distinct from and unconstrained by any public-interest test. Cross-linked the new note from Utgivningsbevis and people-search services and Naming convicted people in Swedish media as a comparator outside Swedish law.
  • Added a collateral-harm section to Unmasking pseudonymous public figures covering people swept into Banksy speculation who were not the eventual subject of the Reuters report: Robert Del Naja’s repeated public denials, Neil Buchanan’s 2020 forced website statement, and Robin Gunningham’s own apparent legal name change after the 2008 Mail on Sunday identification, eighteen years before Reuters called the question settled. The section separates two distinct harms an accuracy-only framing collapses together: harm to a person wrongly named, and harm to a person correctly named without consent, the latter grounded in Banksy’s lawyer’s statement that the artist had “been subjected to fixated, threatening and extremist behaviour” for years.
  • Extended Unmasking pseudonymous public figures from a user-supplied YouTube video (Brew, “How Banksy Was Doxxed”) that recounts 26 years of failed identification attempts while deliberately withholding the name itself. Added two independently corroborated cases: a 2016 Queen Mary University of London geographic-profiling study that named the same person as the 2008 tabloid report, where Banksy’s legal team delayed publication over the analysis’s own privacy-breach risk to that named individual; and George Georgiou, a maintenance man wrongly pictured as the newly unmasked Banksy by outlets illustrating the March 2026 Reuters story, who received up to thirty calls a day and had to be publicly corrected by Reuters itself. Georgiou’s case is analytically distinct from Del Naja’s and Buchanan’s: it was produced by the publishing industry’s demand for an image to run beside an already-published name, showing that resolving the underlying identity question does not end the collateral-misidentification risk — it can relocate that risk onto a new, unconnected person.
  • Added a Scott Alexander (Slate Star Codex) comparator section to Unmasking pseudonymous public figures after the user named the case directly. The 2020 New York Times threat to publish his surname, and his subsequent self-disclosure under the Astral Codex Ten name before the Times profile ran, isolates the same publisher-versus-subject structure from any question of illegal underlying conduct: Alexander’s anonymity protected a psychiatrist-patient relationship and personal safety against documented past threats, not legal jeopardy from the blogging itself. Recorded the Times’s inconsistent pseudonymity standard (contrasted with its 2020 Chapo Trap House profile) and identified pre-emptive self-disclosure as a third unmasking-resolution route alongside the Banksy case’s third-party investigative disclosure and state legal process — one available to Alexander specifically because naming himself carried no criminal exposure, unlike Banksy. Updated the note’s “why the case matters” section to state explicitly that the recurring pattern does not depend on the underlying activity being unlawful.
  • Ran the deep-research skill’s multi-channel workflow against the Banksy and Scott Alexander cases after the earlier passes used plain web search only, and renamed the resulting note to Unmasking pseudonymous public figures now that it develops three co-equal cases rather than one anchor case with comparators. Exa’s research paper and general web search surfaced material no prior pass had found: a Völkerrechtsblog legal analysis applying the ECtHR’s Von Hannover/Axel Springer proportionality test and Halet v Luxembourg compelled-identification doctrine to Banksy, including the previously missing detail that the September 2025 Royal Courts of Justice mural responded to the UK’s Palestine Action proscription under the Terrorism Act 2000; the Berne Convention’s Article 7(3) right to anonymous or pseudonymous authorship, connected to Banksy’s trademark-over-copyright strategy in Full Colour Black v Pest Control; and art historian Peter Bengtsen’s own 2016 methodological critique of the geographic-profiling study, identifying him as the same recurring critical voice behind the 2026 Conversation piece already cited. Added a third full case, the 2014 Newsweek identification of Dorian Nakamoto as Bitcoin’s pseudonymous creator — a confirmed misidentification with no equivalent to Reuters’s handwritten confession, imposed on a retired engineer recovering from a stroke — and used David Douglas’s scholarly taxonomy of doxing (deanonymizing, targeting, delegitimizing) to place all three cases inside an existing academic frame. Extended the Scott Alexander section with Cade Metz’s own retrospective: after Alexander’s farewell post went viral, his supporters leaked Metz’s editor’s name, address, and phone number, showing the dispute produced collateral harassment running in both directions independent of who was right about the original public-interest question. Added Reuters reporter Simon Gardner’s own strongest defense of the investigation from an NPR interview, including his account of deliberately withholding Banksy’s region and further identifying detail.
  • Audited Index navigation and added two top-level topic entries. The pornography-literacy domain (34 notes, five open questions) reached the Index only through one incidental link two hops down under the age-verification topic, so added an entry pointing at Pornography literacy and harm reduction. The Swedish public-record and transparency-privacy domain (people-search services, naming convicted people, protected personal data, pseudonymous unmasking) was reachable only through mid-level notes and no listed hub, so added an entry pointing at its synthesis hub Swedish public-record privacy. Both target existing strong hubs, so no new index page was warranted.
  • Synced the Graphs Curated overview story graph to that Index change. The small map is the counterpart to the landing page, and it still showed only five top-level routes, so it had become materially stale. Added synthesis nodes for Pornography literacy and harm reduction and Swedish public-record privacy as new routes from Index, keeping the single existing venture-to-tax overlap rather than inventing a hub-to-hub edge that does not exist. Verified in a served browser that all five Mermaid blocks render to SVG with no error markers, the two new nodes navigate to their static-site URLs, and the complete force graph initializes at 329 nodes and 1256 links.

2026-07-22

  • Retyped four notes from concept to synthesis after a review of whether declared type matched each note’s warranty. The distinction that decided them: a concept note defines a thing (“X is such-and-such”), while a synthesis note asserts a proposition and argues it from its sources toward a conclusion the vault has a stake in. Retyped Unmasking pseudonymous public figures (a cross-case thesis argued from Banksy, Scott Alexander, and Nakamoto), False accusations as an epistemic risk (the false-accusation rate is latent, so prevalence arguments are circular), Witness reports are not ground truth (a witness is a lossy measurement process, argued from the eyewitness literature), and Probabilistic interpretations of beyond reasonable doubt (a worked derivation from the “98 percent” gloss to its one-in-fifty consequence).
  • Left four thesis-titled notes as concept on the same test: Data as coercive power, Confidentiality as safety infrastructure, Ethics of resisting lawful extraction, and Naming convicted people in Swedish media each read as definitional building blocks that synthesis notes reference and argue from rather than arguments in their own right — Case for privacy and security, for one, states its core claim as privacy limiting the conversion of information into Data as coercive power. That link structure, not the thesis-shaped titles, settled the calls.
  • Added a “Type warranty” check to the audit-structure skill’s per-cluster reading pass, since note type was previously unaudited: neither just links, just clusters, nor the skill ever tested whether a declared type matched what the note does. The check carries the opening-move signal and the link-structure disambiguator that resolved this session’s calls. Ran just sync-skills.
  • Corrected Less harmful social media after review found its “Harassment and unwanted adult contact” table row invented an “adult” qualifier absent from the cited DSA protection of minors guidelines taxonomy, which separates conduct risks (cyberbullying) from contact risks (grooming, coercion, sexual extortion) without an adult-perpetrator qualifier on either. Split the row into peer harassment and grooming/adult-contact to match the source. The deeper problem was that the note, despite framing itself as the alternative to platform exclusion, only developed contact-restriction and architecture responses and gave education, disclosure, and trusted-adult approaches a single dismissive mention. Ran the deep-research skill and added two new sections with comparable evidentiary depth to Pornography literacy and harm reduction: a Spanish RCT found brief grooming-recognition education reduced sexualized responses to actual solicitation, and a Cochrane review found school-based prevention programs improve protective knowledge and behavior with no measured harm, though the disclosure-rate effect loses significance after a clustering correction most included trials originally omitted. Population evidence (pan-European, Finnish, and US) shows adolescents overwhelmingly disclose to peers first, not adults or platforms, which reframes contact-restriction-only design agendas as optimizing for the disclosure path used least. Peer bystander training has direct but mixed trial evidence: a Belgian cluster-RCT of the Friendly Attac game improved bystander self-efficacy, a Norwegian cluster-RCT of the NettOpp app found no significant effect, and a Swedish study tied defender behavior to student-teacher relationship quality. On the user’s displacement hypothesis — that hardening contact controls will just make remaining offenders improve their opsec — the evidence is real but narrower than assumed: a Protect Children survey of 30,000 self-reported offenders and IWF reporting document offenders migrating to end-to-end encrypted platforms specifically to evade content detection, but no comparable evidence was found for displacement driven by contact-restriction features, so the note records that distinction rather than treating the analogy as demonstrated. Added a further structural section distinguishing a federated, volunteer-run network on a shared protocol such as ActivityPub from the note’s existing “different business model” paragraph: removing the entity whose revenue depends on maximizing engagement is a governance change, not merely a different way of funding the same centrally operated design, and it removes a mechanism the JRC review and the Commission’s TikTok and Meta findings identify as central to compulsive use, at the cost of less centrally consistent moderation and safety defaults across independently run servers. No outcome study was found testing this proposal directly, so the section is recorded as structural and mechanistic reasoning rather than as tested evidence.
  • Added a section to Less harmful social media on why adult-minor interaction constraints depend on age verification that current systems cannot deliver. A Knight-Georgetown Institute technical assessment found every deployed age signal has accuracy or availability limits and that no system can block all minors without also blocking many adults, and a 2026 study of Australian adolescents documented a culture of identifying and sharing verification-bypass techniques. The section argues, from routine activity theory’s guardianship logic, that a leaky age boundary fails asymmetrically: non-motivated adults have little reason to defeat the check, while a motivated groomer has a strong reason to, so the constraint can remove ordinary adult presence from a minor’s contact pool faster than it removes motivated offenders. Flagged this explicitly as an assembled mechanism, not a demonstrated result, and included a directly discouraging counter-finding: one existing study found guardianship-improving measures were the least effective routine-activity factor against general online victimization.
  • Corrected a citation error in that same section: the Australian children’s-circumvention study had been cited as evidence for adults defeating an adult-minor contact filter, but the study is about minors evading a platform-wide age ban, not about anyone bypassing a contact restriction. Removed the misapplied citation and its now-orphaned footnote rather than patching around it.
  • Added a further section to Less harmful social media distinguishing minor-only enclaves (a space that excludes adults entirely, as on Discord-style servers) from the adult-minor contact-filter case just corrected. The same asymmetric-incentive logic applies with the roles reversed: a legitimate adult (parent, teacher, youth worker) has little reason to evade a minor-only space’s entry check, while a motivated groomer has every reason to, including posing as a minor, which documented Discord grooming cases show offenders doing with voice-altering software. The academically documented “764” extortion network, which operated across youth-heavy Discord, Minecraft, and Roblox servers with little effective adult oversight, is cited as one documented case of the mechanism, not a base-rate estimate. General protective-factor evidence (authority-figure absence as an independent risk factor for adolescent deviance; social isolation and distrust of institutions as recurring features of commercial-sexual-exploitation contexts) supports the underlying logic without testing the enclave case directly, and the section says so explicitly. Added a callout distinguishing adult presence from adult accountability, since a vetted, trained, paid-moderator minor-only service is not the same failure mode as one with no accountable adults at all.
  • Rewrote the adult-minor contact-restriction callout in Less harmful social media after the user flagged “this does not argue against restricting adult-minor contact” as exactly the reflexive hedge AGENTS.md’s calibrated-qualification guidance warns against: a genuine conclusion stated, then walked back at the moment it should land. The section’s own evidence already supported the stronger claim — most adults pose no risk and have no reason to evade a rule that costs them nothing to obey, while a motivated offender has every reason to defeat it, so a blanket restriction selects the benign majority out of a minor’s reachable adult population faster than it selects offenders out — so the callout now argues that conclusion directly and prefers targeted interventions (education, disclosure, trusted relationships, narrower contact-pattern-specific restrictions) over a blanket restriction as the default response, rather than merely pairing them. Retitled the section heading to match (“selects against its own purpose”). The surrounding causal-chain paragraph keeps its existing honesty standard: the mechanism is evidenced and coherent, the enclave-level outcome has not been directly tested, and the policy conclusion follows from the mechanism under that acknowledged uncertainty rather than from a claim that the outcome is already proven.
  • Corrected the same callout’s example of a “narrower, evidence-targeted restriction” after the user pointed out it had smuggled stranger-danger framing back in: “limiting first contact from an unconnected adult stranger” does not follow from the note’s own asymmetric-incentive argument, which distinguishes motivated offenders from everyone else, not strangers from acquaintances. Most sexual abuse of minors is committed by someone the child already knows, not a stranger, which makes “stranger” a doubly wrong target category. Rewrote the callout to say so explicitly, cited McAlinden on grooming and offender familiarity, and added the natural-mentoring meta-analytic literature (van Dam et al.) showing nonparental, non-relative adults deliver measured protective benefit to adolescents, not merely neutral benefit, to ground “it takes a village” in actual evidence rather than using it as an unsupported rhetorical gesture. Replaced the stranger example with a risk-pattern example instead (rapid escalation to secrecy or sexual content, repeated contact after a minor disengages).
  • Added two general-purpose checks to AGENTS.md after this session’s Less harmful social media edits produced two related failures that existing guidance did not fully cover. The Analytical Voice section gained a check for narrowed claims: when prose moves from a general claim to a targeted one, verify the narrower category is what the cited evidence supports rather than a culturally familiar proxy for it (the session’s own failure: substituting “unconnected adult stranger” for “motivated offender” without evidence for that substitution). This generalizes the existing sexuality-section substitution check beyond sexuality questions, without moving or duplicating that section’s own worked examples. The Sources and Citations section gained a citation re-verification check: re-read a citation’s attached sentence after edits that change what the sentence argues, since a source found for one claim can survive a rewrite that makes it support a different, unverified claim (the session’s other failure: a study about minors evading a platform ban, kept attached to a claim about adults evading a contact filter after the surrounding argument shifted).
  • Preserved Ken LaCorte’s YouTube video “Should innocent people talk to the police?” as a raw/rendered HTML pair and a plain-text transcript, and added Should innocent people talk to the police as its source note. The video argues, from a US-centric legal frame, that innocent people can be harmed by voluntary police cooperation through evidentiary asymmetry for exculpatory statements, stress-driven memory and behavior effects, the undisclosed witness-to-suspect status shift, and silence itself being usable as evidence under Salinas v. Texas. The timestamped-transcript tool was IP-blocked by YouTube at import time, so the source note records that gap and stays working until a timestamped transcript can replace the plain one. Completed the placement pass this import initially skipped: added Coercive interrogation and false confessions as a new concept note comparing the video’s Norfolk Four case to the Bergwall commission material already grounded in Rättssäkerhet in Swedish criminal cases, since both describe the same coerced-false-confession mechanism operating through different pressure tactics in different legal systems. Cross-linked the three notes. Left “Witness reports are not ground truth” unlinked: its subject is third-party perception and memory error, a distinct mechanism from a suspect’s own stress-degraded self-presentation under interrogation.
  • Added a mandatory placement pass to the import-source skill and AGENTS.md, and a rule that a bare URL with no other instruction is always an import-source request rather than a question to ask the user. Also added YouTube-specific guidance to fetch the timestamped transcript rather than the plain transcript, since only a timestamp supports claim-level citation. This followed a session where the earlier LaCorte import stopped at the source note without searching the vault for connections, and used the plain transcript by default.
  • Imported James Duane’s 2008 lecture “Don’t Talk to the Police” (Regent University School of Law, 47 minutes, YouTube upload 2012-03-20), the primary source Should innocent people talk to the police credits as changing LaCorte’s own view. Saved the raw/rendered HTML pair and, this time, the timestamped transcript, and added Don’t talk to the police as its source note with timestamp-anchored claims throughout. The lecture argues that even innocent people should never talk to police, through hearsay exclusion of exculpatory statements, an Innocence Project statistic that over a quarter of DNA exonerees made incriminating statements, Ohio v. Reiner and Ullmann v. United States on the Fifth Amendment, a live memory-contamination demonstration, and a co-presenting Virginia Beach detective’s corroborating account of interrogation tactics. Completed the placement pass: linked the LaCorte source note to this one where it names Duane, and strengthened Coercive interrogation and false confessions with Duane’s Eddie Joe Lloyd and Earl Washington cases, which show the same mechanism operating through information contamination exploiting mental illness and cognitive impairment, a third pattern alongside Norfolk Four’s threats and Bergwall’s suggestive therapy.
  • Imported two files the user provided directly: Duane’s 2010 ABA article “The Right to Remain Silent, A New Answer to an Old Question” (Criminal Justice, SSRN 1998119) and his 2016 book You Have the Right to Remain Innocent (Little A). Preserved both under src/sources/, dated by original publication year, and added The right to remain silent and You Have the Right to Remain Innocent as their source notes. The book expands the lecture’s thesis with an overcriminalization argument and a wider wrongful-conviction case catalog, then adds two parts absent from the lecture: “Don’t Plead the Fifth,” arguing that Salinas v. Texas lets plain silence be used as evidence of guilt while United States v. Long shows that naming the Fifth Amendment explicitly is not safe either, and “Plead the Sixth,” recommending an unambiguous request for counsel instead, since courts have not yet treated that the same way. Completed the placement pass: added Invoking the Fifth Amendment can itself be used as evidence as a new concept note synthesizing this argument across all three Duane sources, linked it from Should innocent people talk to the police where that video separately cites Salinas v. Texas, and strengthened Coercive interrogation and false confessions with the book’s juvenile false-confession statistic. While drafting the new concept note, caught and corrected an attribution error before it lint-passed: a claim that Duane’s lecture cites Salinas v. Texas actually belonged to LaCorte’s video, which cites Duane as its inspiration but argues that specific point independently.
  • Ran a deep, independent research pass on the “don’t talk to police” claim cluster, since every note on the topic so far rested on Duane’s own advocacy and LaCorte’s summary of it. Three parallel research threads verified Duane’s case citations against primary legal sources, checked his statistics against the peer-reviewed false-confession literature, and researched counterarguments and the current Miranda doctrine trajectory. Findings required real corrections, not just additional citations: Salinas v. Texas is a fractured 3-2-4 decision, not the clean 5-4 precedent it is usually described as, and its holding presupposes prior police questioning had already occurred, not spontaneous silence; United States v. Long never actually resolved the constitutional question Duane cites it for, across two separate appellate opinions; Ohio v. Reiner arose in the narrower context of a witness’s immunity grant, not a general defendant’s-rights rule; and Ullmann v. United States’s “shelter for wrongdoers” line is oft-quoted dicta, not the case’s holding. Corrected all four in Invoking the Fifth Amendment can itself be used as evidence, You Have the Right to Remain Innocent, and The right to remain silent. Also found and fixed a citation error already in the vault: Coercive interrogation and false confessions attributed both a juvenile/adult false-confession rate and an average interrogation length to one undifferentiated “340-case study”; these are two distinct sources (Gross, Jacoby, Matheson, Montgomery, and Patil 2005 for the former; Drizin and Leo 2004 for the latter), now cited separately. Added Why you shouldn’t talk to the police as a synthesis note grounding the whole topic in verified law and the empirical literature. Its central findings: Duane’s practical advice holds up and has become more important, not less, since Vega v. Tekoh (2022) removed the civil remedy for Miranda violations; the strongest critique of his position is a documented selection-bias argument from independent commentators (his examples are drawn only from people who talked and then needed a defense attorney), not a legal error; Duane has quietly narrowed his own public position since the original lecture; a 2024 systematic review found non-accusatorial interviewing methods produce both fewer false confessions and more true confessions than the Reid technique; and England and Wales’s 1994 reform of adverse-inference-from- silence rules makes current US law comparatively more protective of silence, not less, complicating any assumption that strong silence protection is an obvious universal norm. Added Does voluntary cooperation with police measurably help innocent suspects as a question note recording a real evidence-base asymmetry the research surfaced but could not resolve: documented harms of talking to police are well-counted, but no comparable literature quantifies how often cooperation helps an innocent suspect. Reassessed Coercive interrogation and false confessions from working to stable, since its remaining limitation (no reliable population base rate for false confessions) is now explicitly stated as a permanent scope boundary rather than a known gap awaiting future research.
  • Extended the police-silence research into Swedish law and into a specific strategic question the user raised: whether it can be rational to accept a known, bounded process-crime sentence — refusing to disclose an encryption key, contempt, obstruction — over risking a shorter but far more damaging conviction built in part on one’s own words or device contents, particularly where the worse charge carries permanent reputational destruction a bounded sentence does not. Added Swedish right to silence and förklaringsbörda to show Sweden’s right to silence is a composite of ECHR Article 6 case law and EU Directive 2016/343, not a codified guarantee; that RB 35:4 is facially more permissive than England’s 1994 reform but has been narrowed in practice by Högsta domstolen’s own förklaringsbörda doctrine, imported nearly verbatim from John Murray v. United Kingdom; that JO explicitly flagged this as potentially incompatible with the Directive during its implementation and was overruled by the government, an unresolved institutional disagreement rather than a settled point; and that RB 23:12 categorically bars the police-deception tactics Frazier v. Cupp permits in the United States. Added Swedish encryption and passcode disclosure law, anchored on a 2023 JO decision holding that a suspect’s PIN is protected by the self-incrimination privilege; Sweden compels only biometric unlock, classified as a bodily act, and has no equivalent of the UK’s Section 49 notice regime. Substantially advanced the existing open question How does Swedish law treat duress wipes with the specific statute a prosecutor would most plausibly plead, BrB 17:11 (skyddande av brottsling), and Högsta domstolen case law (NJA 2019 s. 294, NJA 1990 s. 175) suggesting it likely does not reach a suspect’s destruction of their own device, since it requires protecting another person’s crime; this remains an inference from statutory structure rather than a directly confirmed holding, so the question stays open. Added Encryption key disclosure compulsion comparing the UK’s bounded, fixed-maximum Section 49/53 offense (Oliver Drage and “JFL” both received sentences far under the statutory ceiling, with no substantive charge ever following) against the US’s unsettled foregone-conclusion doctrine and its often-unbounded civil-contempt alternative — Francis Rawls was held over four years, nearly three times the eventual court-recognized cap, before release, and was never charged with the underlying offense. Added Why a known harsher penalty can beat an uncertain lesser one as a synthesis note evaluating the user’s proposition directly: it has genuine decision-theoretic grounding (Segal and Stein’s “asymmetric ambiguity aversion”), real cases where the predicted pattern held (the UK cases), and a decisive real case where the same premise catastrophically failed (Rawls’s civil contempt), because whether a “bounded lesser charge” is actually bounded depends on the jurisdiction’s specific legal mechanism, not on the defendant’s preference. No case was found, in any jurisdiction, of a defendant explicitly stating this exact reputational trade-off as their reason for refusing disclosure. Linked the new material into Duress credentials and coercive extraction, Ethics of resisting lawful extraction, and Rättssäkerhet in Swedish criminal cases.
  • Corrected Why a known harsher penalty can beat an uncertain lesser one after the user clarified their original proposition: it compared a KNOWN five-year sentence for refusing to disclose an encryption key against a KNOWN, numerically shorter three-year sentence for a possibly false CSA charge, with the preference for non-disclosure holding even though the bounded sentence is nominally the longer one — a stronger and more specific claim than the note’s earlier “nominally worse-looking” framing captured. The user also named two further, distinct reasons the note had not covered, beyond collateral consequences: a dignitary reason (conviction through manipulated words versus punishment for principled refusal) and a legal-but-compromising disclosure risk (cooperation surfacing true, legal, but separately damaging material unrelated to the charged offense). Research grounded both. The dignitary reason has real philosophical support: Jennifer Lackey’s concept of agential testimonial injustice names the specific wrong of a conviction built on subverted epistemic agency, complemented by Duff’s communicative theory of punishment and civil-disobedience literature on the intrinsic value of accepting known consequences for principled refusal; the empirical psychology of this specific comparison is honestly thin, and the note says so rather than manufacturing support. The disclosure-risk reason is grounded in three independently verified mechanisms: US Federal Rules of Evidence 413/414 allow propensity evidence from an uncharged, unconvicted prior sexual offense on a mere preponderance standard, though courts have repeatedly rejected merely legal, stigmatizing material for this purpose, detailed in the new Propensity evidence from legal but suggestive material; UK non-conviction intelligence retention (Police National Database “soft intelligence”) is documented official policy; and Sir Cliff Richard v BBC [2018] EWHC 1837 (Ch) is a fully litigated case establishing severe reputational harm from investigation alone, never charged. Re-examining the vault’s existing Rawls and Davis case law found a direct correction: Rawls’s own case record shows partial, voluntary cooperation surfaced material unrelated to the original investigation’s predicate, and Commonwealth v. Davis’s holding was reasoned expressly around this exact scope-creep risk, not merely the compulsion-mechanism question the existing Encryption key disclosure compulsion previously covered; updated that note accordingly. Added Non-offending attraction to minors to ground the user’s NOMAP false-accusation test case in the actual clinical and terminological literature (Dunkelfeld’s zero-reoffense finding among non-offending treatment completers; contested MAP/NOMAP terminology per Farmer, Salter, and Woodlock’s 2024 critique), keeping attraction, offending, and false accusation categorically distinct per the vault’s sexuality-and-age-sensitive-analysis rules, and explicitly stating that no case was found combining a false CSA accusation with disclosed non-offending support-seeking in this exact configuration — the scenario is a plausible, mechanism-grounded construction from independently documented risks, not an observed pattern.
  • Renamed Why a known harsher penalty can beat an uncertain lesser one after the user pointed out its previous title, “Whether accepting a bounded lesser charge beats an uncertain worse one,” implied the process crime carries the lighter penalty — the opposite of the proposition it evaluates, where the known option is the numerically harsher one. The old title is preserved as an alias.
  • Removed process-narration language the user flagged as leaking into article prose across several notes from this research arc and a few pre-existing ones found in the same sweep: phrases such as “the sole focus of an earlier version of this note,” “restated at full strength,” “now separates and cites them correctly,” and “during this research pass” describe the editing session rather than the subject, and are irrelevant to a reader. Corrected Why a known harsher penalty can beat an uncertain lesser one, Why you shouldn’t talk to the police, You Have the Right to Remain Innocent, Swedish right to silence and förklaringsbörda, How does Swedish law treat duress wipes, Jaeger and Kumar group chat encryption analysis, Witness reports are not ground truth, and VPN service legal risk. A note’s own cross-references to its own sections (“this note treats,” “this note develops”) are a different, acceptable pattern and were left alone.

2026-07-23

  • Imported three sources on age verification and youth online-safety policy: the Internet Matters report The Online Safety Act: Are children safer online? (May 2026), preserved as the report PDF; Fortune’s report on children using makeup and borrowed logins to pass age checks; and the Associated Press dispatch on France’s under-15 social-media ban, the latter two preserved as raw-plus-rendered HTML pairs.
  • New source notes: Internet Matters Online Safety Act report, Fortune Gen Alpha age-verification report, and Associated Press report on French under-15 social media ban.
  • New entity notes UK Online Safety Act and French social media ban for under-15s, plus the question note Do under-16 social-media bans reduce measured harm, which tracks the natural experiments now running in Australia, France, and the United Kingdom.
  • The Internet Matters findings propagated into six notes: Age assurance (32% of UK children bypassed a check within two months, mostly by low-tech means), Age-verification circumvention services (everyday circumvention needs no vendor), Virtual private networks and child protection EPRS (children’s VPN use stable near 8%, complicating the spike narrative), Does age verification reduce harms other measures cannot (first post-enforcement snapshot, cross-sectional only), Less harmful social media (blanket bans moving from proposal to law; children name engagement mechanics as their top concern), and Which social-media designs reduce harm for young users (demand-side evidence for feature restrictions).
  • Left Ofcom and Australian under-16 social media ban as redlinks.
  • Fixed a pre-existing lint defect in Todo.md: a literal section-wikilink token in one task failed the unresolved-section check; reworded to name local and page-plus-section links in prose.
  • New entity note Cellebrite covering: company background and product capabilities (UFED, Premium, Inseyets); zero-day acquisition practices including evidence from leaked internal files, the public jailbreak community, and the Corellium acquisition; the damage from hoarding exploits — data breaches (2017, 2022, 2023, 2025), the RAND 2017 study’s low-collision-rate argument and its limitations, authoritarian misuse (Serbia, Belarus, Russia, Bangladesh), and the 2021 Signal demonstration that Cellebrite’s own software is exploitable; defences including GrapheneOS, auto-reboot, USB restrictions, Lockdown Mode, and the hardware asymmetry that leaves most Android devices vulnerable. Added wikilinks to the existing mentions in Mobile-device extraction and evidentiary selection and Rättssäkerhet in Swedish criminal cases.
  • Expanded Pegasus spyware from a 38-line seed to a developed entity note covering NSO Group’s corporate structure and ownership history (Francisco Partners, Novalpina Capital, the 2023 lender foreclosure transferring control to founder Omri Lavie), technical capabilities (zero-click vectors including Trident, Heaven, FORCEDENTRY, and the PATN anonymising infrastructure), the documented abuse geography (El Salvador/El Faro, Morocco, Poland, Hungary, Saudi Arabia/Khashoggi, and the 2021 Pegasus Project investigation of 50,000 phone numbers), legal consequences (US Entity List sanctions 2021, WhatsApp v. NSO Group damages ruling December 2024, Apple lawsuit), and defences (MVT, iMazing Spyware Analyzer, Lockdown Mode, regular reboot). Added relationship to Cellebrite and the broader mercenary-spyware ecosystem.
  • Triaged the wanted-page queue (20 targets). Created 9 new notes: EncroChat (encrypted phone platform shut down in 2020, largest example of unexaminable evidence in Swedish prosecutions), Australian under-16 social media ban (in effect December 2025, $49.5M AUD penalties, first blanket age restriction), Eustella (EU AI startup using Qwen 3.5 and other open-weight models on EU infrastructure, targeting 100M+ European users), Reproducible builds (deterministic compilation closing the source-to-binary correspondence gap), Ofcom (UK communications regulator enforcing the OSA), Ethereum, Solana, and Zcash (cryptocurrency entity notes alongside existing Monero), and PGP and OpenPGP (the existing open email-encryption standard with well-documented failure modes). Removed six generic abbreviation wikilinks (APK, API, SDK, CI, TLS, TLS) that served only as display pipes; the abbreviations are defined locally in frontmatter. Fixed an escaped-pipe wikilink in Secure and decentralized communication stacks’s messenger table. Resolved Ethereum, Solana, and Zcash wikilinks in Monero’s prose. Kept seven redlinks as purposeful: Confluence model of sexual aggression, JMAP, Samir Sabri, Joy Rahman, Esa Teittinen, Android Package, and F-Droid.
  • Completed the triage pass by resolving the remaining seven redlinks: F-Droid (entity, free-software Android repository, inspired by Debian model, per-app signing keys, reproducible-builds transition), JMAP (concept, IETF standard RFC 8620/8621, JSON-based mail synchronisation, Fastmail-developed), Android Package (concept, APK file format, signature continuity without source correspondence), Confluence model of sexual aggression (concept, Malamuth’s hostile-masculinity and impersonal-sex pathways), Samir Sabri (entity, convicted 1986 at age 15, acquitted 2016, SEK 1.2 million compensation 2021), Joy Rahman (entity, convicted 1994, acquitted 2002, 8 years imprisoned, SEK 10.2 million compensation), and Esa Teittinen (entity, convicted 2010, acquitted 2018, 8 years imprisoned). The wanted-page queue is now empty.
  • Added Ontology redesign proposals (meta, working), recording the 2026-07-23 design session’s three unsettled proposals: demand-driven extraction of concepts and entities to serve a planned synthesis (the ledger model), an occasional distill-syntheses skill with decompose and ratchet modes, and the type-to-kind swap with type returning as a per-kind structural vocabulary (entity types: person, organization, product, event, statute). The note also records what was rejected or deferred (flat types, a hard extraction gate, quote, deferred synthesis types, a pilot-conditional data kind) and carries the seven open decisions as tasks.
  • Clarified the note-type definitions in AGENTS.md (spec change, not a proposal adoption): source, entity, concept, synthesis, and index now have explicit definitions, with the concept/synthesis boundary stated as the warranty test ported from the audit-structure skill, and entity’s observed scope documented (persons, organizations, products, and events such as breaches, cases, statutes, and regulatory actions).
  • Reviewed Ontology redesign proposals and folded the review’s findings into the note: the existing syntheses carry note-level provenance only (frontmatter sources with few or no claim-level citations), so the ratchet pilot is a claim-mapping pass before it is an extraction pass; the type-to-kind swap must also sweep AGENTS.md, the skills, and agent memory, where stale type= references fail silently; statute needs a widen-or-accept decision for EU regulations and directives; leave-alone decisions in the ratchet are recorded in Log.md.

2026-07-24

  • Created five new type: meta notes documenting wiki architecture: Wiki type system, Note lifecycle, Linking and citation conventions, Source and evidence model, and Build and rendering pipeline.
  • Created Wiki architecture as a type: index note gathering all meta notes into one entry point.
  • Updated Index to link to Wiki architecture from the Wiki section.
  • Checked off the “Document the wiki architecture and design choices” task in Todo.
  • Imported three YouTube videos on gender and education provided as bare URLs: War on Boys (PragerU, 2014, Christina Hoff Sommers), and from Elephants in Rooms with Ken LaCorte: What if boys were never the problem (July 2026) and Some cold facts about math and gender (May 2025). Saved timestamped transcripts as War on Boys transcript, What if boys transcript, and Math and gender transcript.
  • Created Greater male variability as a type: concept note capturing the biological hypothesis that appears in two of the three sources (males at the tails, X-chromosome mechanism). The three source notes are cross-linked; the new concept note is linked from the two sources it draws on.
  • Synthesis pass: these sources do not materially bear on any existing vault note. The vault’s content is organized around privacy, digital rights, Swedish law, and cryptocurrency; gender differences in education are a genuinely new topic area with no existing subject notes to update. The sources and concept note stand as a new cluster that future imports or research may connect to the vault’s existing work on intergenerational dynamics or adolescent development.
  • Imported Farrell (1993) The Myth of Male Power and Gilbert (2007) The Disposable Male as EPUB artifacts from the user’s Hyacinthus library. Both books develop the “male disposability” thesis — Farrell through a political-economic lens (workplace death, suicide, criminal justice, family court), Gilbert through an evolutionary-psychology lens (greater male variability as a reproductive strategy). Source notes created as The Myth of Male Power and The Disposable Male. Updated Greater male variability to cite both books, and added cross-links from the three video source notes to the two new book notes. The gender-and-education cluster now spans five sources plus one concept note.
  • Deep-researched the intersection of all five gender-and-education sources against the external scholarly literature. Key findings: greater male variability is well-established (Hedges and Nowell 1995, Feingold 1992, Gray et al. 2019) but its interpretation is disputed — Hyde’s 2010 meta-analysis calls VR = 1.08 “nearly equal,” while Hill and Arden (2023) showed this inference is statistically unsound without distributional assumptions. The boys’ grade penalty is real (Voyer and Voyer 2014; d = 0.225) and partly mediated by teacher bias (Terrier 2020: 5.2% bias in math grading; Lavy and Sand 2018), but mostly driven by non-cognitive skill differences (Cornwell, Mustard, and Van Parys 2013). Male teachers do not improve boys’ outcomes (Helbig 2012, 21 countries). The ADHD relative-age effect is well-replicated (Sayal et al. 2017: youngest-in-class 26–31% more likely diagnosed). The education–suicide link is confounded rather than causal (Lorant, Kapadia, and Perelman 2021), complicating LaCorte’s downstream-consequences argument. The gender-equality paradox (Stoet and Geary 2018) is partly contested as a measurement artifact (Richardson, Reiches, Bruch, et al. 2020) but the broader finding — more sex differences are larger, not smaller, in countries with better living conditions — withstands criticism (Herlitz, Honig, Hedebrant, and Asperholm 2024). Updated Greater male variability from seed to working with scholarly evidence and the interpretation dispute. Created Male variability, education, and life outcomes as a type: synthesis note mapping agreements (grades gap real, GMV real, the real problem is the bottom tail) and disagreements (hostile schools versus developmental mismatch, policy fixability, structural versus biological causation) with claim-level external citations. Cross-linked the synthesis from all five source notes and the concept note.
  • Imported seven source artifacts under src/sources/ covering Qubes OS architecture and TCB documentation, secureblue features and FAQ, Kicksecure about and comparison pages, and the Kicksecure vs secureblue development comparison.
  • Created four source notes: Qubes OS architecture documentation, secureblue features and FAQ, Kicksecure security features documentation, and Kicksecure vs secureblue development comparison.
  • Created three entity notes: Qubes OS, secureblue, and Kicksecure, covering each project’s security model, hardware requirements, deployment options, and relationships to one another and to Whonix.
  • Created synthesis note Desktop operating system security comparison comparing the three across isolation architecture, hardening depth, hardware tradeoffs, anonymity integration, and complementary deployment patterns.
  • Deep-researched clear backpack policies in United States K-12 schools. Preserved four source artifacts: ACLU Digital Dystopia report (PDF), Secret Service NTAC Protecting America’s Schools (PDF), Patch clear backpacks debate overview (HTML), and CNN Marjory Stoneman Douglas scraps clear backpacks (HTML). Two additional load-bearing sources cited by URL: Nance (2017) Student Surveillance, Racial Inequalities, and Implicit Racial Bias and Stilwell et al. (2025) Securing schools, protecting minds scoping review.
  • Created concept note Clear backpack policies describing the policy, its history, spread, costs, and exceptions.
  • Created synthesis note Clear backpacks are school security theater concluding that no evidence supports the policy, the attack vectors defeat the mechanism, the harms are documented and unequal, and the evidence supports threat-assessment alternatives instead.
  • Created concept note Behavioral threat assessment in schools describing the NTAC, CSTAG, and Salem-Keizer models and the causal evidence for their violence-prevention effects.
  • Created concept note School surveillance and the school-to-prison pipeline documenting the adoption disparity, pipeline mechanisms, and why clear bags belong in the bundle.
  • Created question note What interventions reduce school weapon carriage recording the thinness of the evidence base for the outcome the debatest turns on.
  • Left redlinks for Columbine High School massacre, Marjory Stoneman Douglas High School shooting, Comprehensive School Threat Assessment Guidelines, and K-12 School Shooting Database.
  • Added two new tags: school-security and school-to-prison-pipeline. Lint and build pass.

2026-07-25

  • Rewrote Todo into sections (architecture, renderer, ontology redesign, agent workflow, Obsidian feature scope, done) and recorded the conclusions of a design session about the wiki’s own architecture, so the reasoning survives without the conversation.
  • Settled and closed four architecture questions in that note: no single-page full-JS rewrite, no port of the build to JS, no server-side rendering for now, and updated stays in frontmatter but should be stamped by just fix. The build’s real ceiling is the eagerly loaded index, not note count or the choice of renderer.
  • Corrected the single-page-app reasoning twice during the session, and reopened the question. A fully inlined app reads from file:// as well as the current build does, by the same script-tag trick that already loads wiki-data.js, since CORS blocks fetch against local files but not script elements. Lazy loading per-note files is what forfeits file://, and it does so for a static site just as much as for an app. The second correction was the larger one: scale is decided by where rendering happens, not by how navigation works, so the conclusion that server-side rendering is unnecessary says nothing about whether the client should be an app or prerendered pages with navigation added. That question is now recorded as open, turning on whether a note should stay independently addressable.
  • Measured the index problem rather than estimating it: dist/wiki-data.js is 1.6MB against 1.86MB of source Markdown, because its search field stores each note’s visible prose, so the build’s “index” is nearly a second copy of the vault. Queued replacing it with a chunkable inverted index and a separate metadata-plus-excerpt tier.
  • Queued moving the structural data out of wiki-data.js and into the pages that need it. Only one of its ten published fields is about search; the browser currently rebuilds the whole note graph and the tag co-occurrence graph on every page load from data the build already held and already derives for just links and just clusters. Backlinks, per-note neighborhood graphs, and tag co-occurrence are build-time facts, and baking them in leaves search as the only global payload, which can then load on demand instead of on every page. The blob exists to sidestep build-time invalidation, a fair trade at 408 notes and not at 1.6MB, and the dependency set is computable in the pass that already builds the edge set. Recorded that this pulls against the single-page-app option, since baking structure into pages makes notes self-sufficient while an app wants the global data resident.
  • Queued extracting the CSS and JS out of bin/wiki.py, where they live as string literals inside a roughly 1,900-line layout function, and scoping Playwright as an optional dependency so lint, build, and test stop resolving a browser automation library none of them reach.
  • Recorded a linter bug found while editing the note: masked_body_lines masks fenced code blocks but not inline code spans, so backticked wikilinks, highlights, and section links are linted as real syntax across all four callers. AGENTS.md already states that a code span is a mention rather than a use, but only the instruction-file checks honor it.
  • Rewrote the reasoning against browser-side rendering, which had claimed whole-vault work cannot happen in the browser while the note graph and tag co-occurrence graph are computed client-side over every note on each page load. The demonstration cuts the other way: it shows the cost of keeping the vault resident, which is what the build-time structural data task exists to stop paying. Two of the five items listed as whole-vault work were not: abbreviations come from the note’s own frontmatter, and callouts are line-local. The argument that stands alone is build-time MathML conversion.
  • Separated two questions the session had answered as one. “Port the build to JS” was answered as though it meant moving rendering into the browser, which it did not: a Node build script does the same whole-vault work the Python one does, at the same time. Kept the settled answer under its actual question and queued the language question as open. Its real obstacle is latex2mathml, which has no clean JavaScript equivalent, and its motivation is largely the stylesheet and browser script living inside the Python as string literals, so the asset extraction should come first.
  • Reviewed the Markdown library choice and how the renderer uses it. The extension use is mostly sound: the per-note abbreviation glossary, the preprocessor and treeprocessor registration, and the callout markdown="1" reprocessing all go through the real library API. The exception is decorate_headings, which regexes heading tags out of rendered HTML and pairs them with a separate heading parse by position, where the existing treeprocessor would receive the real elements. Queued that fix, plus a lower-priority one for the Mermaid and Vega fence substitutions. Recorded that replacing python-markdown with markdown-it-py is only worth considering if Obsidian and the wiki are seen to render the same source differently, since that divergence is what its lack of CommonMark compliance would cause.
  • Added scheduling tasks for Ontology redesign proposals: run the extraction pilot before writing any skill, decide the type-to-kind swap early because its sweep cost grows with note count and instruction surfaces, and cost the ratchet pass against re-reading rate rather than note count. The open ontology decisions stay in that note.
  • Replaced the blanket Obsidian-feature exclusion with a narrower rule to record in AGENTS.md: a note must stay readable as Markdown and renderable without the plugin that produced it. Dataview and Excalidraw are declined under it; Canvas, Kanban, and task due dates are queued as open decisions. Lint and build pass.
  • Extracted the stylesheet and the browser script out of bin/wiki.py and into share/assets/, where an editor, a formatter, and a diff can treat them as the languages they are. The file went from 5,569 lines to 4,207, and every rendered page and both emitted assets are byte-identical to the previous build. Nine escape sequences in the search tokenizer had been doubled for Python’s string literal and became single backslashes on the way out; searching övergreppen and overgreppen both return four notes, which exercises the diacritic-folding expression that carried them. node --check now validates the script, which a string literal could never support.
  • Fixed three contained defects in bin/wiki.py. masked_body_lines masked fenced blocks but not inline code spans, so syntax quoted in backticks was linted as real syntax; wanted_links bypassed the masking entirely, so quoted and fenced wikilinks were entering the research queue. page_headings line-scanned the raw body, so a commented line inside a fenced block became a phantom heading and shifted every later heading’s copy link onto the wrong heading, though no note in the vault triggered it. Playwright moved to a download dependency group, taking the default environment from fifteen packages to eleven.
  • Fixed the round-trip failure the extraction work surfaced. Securing schools, protecting minds was not at fault: the YAML writer never set a width, so ruamel’s 80-column default folded that note’s PRISMA-ScR expansion onto a second line and just fix would have reformatted the note on any pass touching it. Setting the width repairs every long frontmatter value at once instead of shortening one expansion and leaving the next long one to hit the same edge. The full suite now passes at 85 of 85, and just fix leaves the vault byte-identical.
  • Resolved a contradiction between two tasks written in the same pass. One recorded that lazy loading forfeits file:// access; another proposed fetching the search payload on demand. Script injection loads a deferred file offline where fetch cannot, so on-demand search survives if the deferred file stays a script assigning to a global. Chunking does not survive: fetching postings for arbitrary typed terms is what script injection cannot do responsively, so a chunked index and offline search are mutually exclusive. Recorded that fork rather than assuming chunking is free, including in the settled server-side-rendering answer, which had claimed a chunked index keeps the offline property. Reading the vault from disk is a property with a scale limit, not one the architecture preserves indefinitely.
  • Replaced black with ruff and wired both its linter and formatter into just lint and just fix. black had been a dev dependency and a config block that no recipe ever invoked. The formatters agree closely enough that adopting ruff cost no churn: the rendered site is byte-identical across the change. The linter immediately found a defect a formatter could not. bin/wiki.py used a parenthesis-free multiple-exception except, which is Python 3.14 syntax, while pyproject.toml declares support from 3.11; the file failed to parse on 3.11, 3.12, and 3.13 and worked only because the tooling pins a managed 3.14. The black config had declared py314, which is how the gap stayed invisible, so the ruff target now tracks requires-python instead. Also collapsed four copies of the same fence-detection expression onto the tuple form startswith already supports.
  • Finished the asset extraction, which the first pass had left half-done. Taking only the page-wide stylesheet and script had reduced layout from about 1,900 lines to 1,033, which is not the outcome the task described: the force-graph, Mermaid, and Vega scripts were still inline string literals inside the function. Extracting those three brings layout to 168 lines and bin/wiki.py from 5,569 to 3,842. What remains in the function is the HTML page template, an f-string interpolating about a dozen Python values, which is structure rather than a foreign language embedded in one. All 408 rendered pages are byte-identical across the change, and the three scripts were verified rendering from file://: a Mermaid diagram to SVG, a Vega chart into its inserted frame, and the force-graph canvas.
  • Researched mobile operating system security and imported two sources: Mayrhofer et al., “The Android Platform Security Model (2023)” (arXiv:1904.05572v3, the current revision the user asked for in place of the 2019 original) as sources/2026-07-25-android-platform-security-model-2023.pdf with source note The Android Platform Security Model, and The Hated One’s 2022 GrapheneOS video as a timestamped transcript with source note The Hated One on GrapheneOS (kept despite its age because its developer-reviewed concepts remain accurate; its dated claims are flagged in the note).
  • Created Mobile operating system security comparison comparing GrapheneOS, stock Pixel, other Androids, and iOS across Cellebrite-class forensics, Pegasus-class spyware, commodity malware, and privacy-setup friendliness. Key findings: leaked Cellebrite matrices make GrapheneOS the only documented platform resisting BFU and AFU extraction; Apple’s Memory Integrity Enforcement (iPhone 17, 2025) is the biggest stock-system move against mercenary spyware and is expected to raise forensic-extraction costs too; hardware tier (Titan M2 / Knox Vault / Secure Enclave vs shared-CPU TEE) predicts forensic resistance better than software choices.
  • Created concept notes Titan M secure element and Memory Integrity Enforcement.
  • Expanded GrapheneOS with its actual hardening layers, Pixel 10 support, and the March 2026 Motorola partnership (first non-Pixel path, preinstalled flagships from 2027); added MIE to Pegasus spyware and Cellebrite; connected the platform security model paper’s sideloading trust-on-first-use boundary and its store-held-signing-key footnote to Android app distribution trust models.
  • Moved every note title from a body H1 into frontmatter title, with an optional subtitle, and migrated all 413 notes. The H1 was redundant on both surfaces, since Obsidian shows the filename above the note and the renderer emitted a second copy from the body, which is why a note displayed its title twice. The old rule was also enforced by lint and consumed by nothing: build_resolver keys on the filename stem and aliases, and Note.title returns the stem, so link resolution never read the H1 at all.
  • Kept the filename as the note identity. Obsidian and Markdown Oxide both resolve wikilinks against it, so links, URLs, backlinks, and graph nodes are unchanged, and title is display text that also feeds the search index so a real title stays findable from a shortened filename.
  • Let just new and just rename derive a Windows-safe stem from a free-form title. Agents had been shortening titles by hand and losing information: subtitles truncated at the colon, commas substituted for colons, em dashes flattened to spaced hyphens, and citation identifiers such as SOU 2024:75 and C-136/17 stripped to SOU 2024 75 and C-136 17. Source notes were worst affected, because they name external works whose titles are not ours to choose.
  • Enforced the new shape with rumdl rather than a custom check. It reads a frontmatter title as a top-level heading, so MD041.level = 2 rejects a body H1 and MD025 rejects carrying both, which makes exactly one title source per note structurally required.
  • Recorded the reasoning in Todo and rewrote the note-format and file-naming rules in AGENTS.md.
  • Researched the offline brute-force floor of Android and iOS disk encryption and created Offline brute-force resistance of phone disk encryption: verified against AOSP and Apple primary documentation that Android stretches the credential with scrypt at ~25 ms / 2 MiB (which AOSP concedes “usually does not provide much security”), that Weaver secrets are released by the secure element on correct-key presentation, that Apple’s ~80 ms key-tangle runs inside the Secure Enclave against a fused UID with no read path, and that iOS evicts Complete Protection class keys seconds after locking while Android keeps CE keys resident for the session. Recorded three calibrations: Quarkslab’s Titan M extraction (CVE-2022-20233) was on the Titan M1, needed AP root, and could not read the key-ladder root; the Weaver secret alone is insufficient (the SoC hardware key is also required); GrapheneOS profile end-session is finer-grained than iOS eviction for secondary profiles.
  • Created the question note Does AI-assisted vulnerability discovery erode mitigation-based security on whether Mythos-class AI exploitation erodes GrapheneOS’s mitigation-based lead faster than defensive uses of the same models, with the Anthropic Mythos disclosure, the training-data-contamination critique, and the AISI calibration as the evidence base.
  • Threaded the findings into Mobile operating system security comparison (new cryptographic-floor section), Titan M secure element (Weaver release property and the Quarkslab precedent), and GrapheneOS (why a strong passphrase is the load-bearing user choice).
  • Retyped Titan M secure element from concept to entity: it is a specific Google product line, not a reusable mechanism; the concept form of the subject (discrete secure element vs shared-CPU TEE) would be a separate page if it is ever written.
  • Extended the mobile-security research to Samsung: created Samsung Knox covering the Knox Vault architecture (dedicated processor, separate secure flash, CC EAL4+, NIAP MDFPP / CSfC government certifications), Samsung Weaver on the Knox Vault Processor (same release-on-correct-key design as Titan M, so the same offline-regime structural property), the Knox warranty fuse, the seven-year update commitment, and the forensic record: Cellebrite Inseyets 10.0 added BFU/AFU/FFS extraction for the Galaxy S24 series on both Qualcomm and Exynos, which makes Samsung the demonstration that secure-element-class hardware alone does not deliver Pixel-class extraction resistance. Noted the coming MTE toggle in One UI 9 and that Samsung devices cannot run GrapheneOS (locked NA bootloaders, irreversible Knox fuse).
  • Updated Offline brute-force resistance of phone disk encryption with Samsung Weaver, Mobile operating system security comparison (Samsung rows, BFU table line, MTE status), Titan M secure element (Knox Vault table row), and Cellebrite (S24 BFU extraction in the hardware-asymmetry section).
  • Answered why Knox Vault devices are Cellebrite-extractable when Titan M2 devices are not: imported Osservatorio Nessuno’s “Demystifying phone unlocking tools” (May 2026) as source note Demystifying phone unlocking tools and added a mechanism section to Samsung Knox. The exact Samsung exploit path is Cellebrite-secret, but the documented components are Quarkslab’s end-to-end TEEGRIS patching PoC, Samsung’s more exposed download/EDL boot modes, slower patch propagation, and Cellebrite’s investment-weighted targeting; whether the S24 capability patches the TEE around Knox Vault or breaks the Vault itself is publicly unknown.
  • Upgraded the Weaver-extraction claim in Offline brute-force resistance of phone disk encryption from inference to Quarkslab’s explicit confirmation that Weaver keys and values were extracted from Titan M flash with their read primitive.
  • Added an “is Titan M2 one exploit away” calibration to Titan M secure element: its resistance is empirical (small attack surface, post-M1 redesign, unclaimed .5M bounty, lower attacker investment), not principled, which is why the passphrase stays the final backstop.
  • Imported Landau, Listening In (Yale, 2017) as sources/2017-landau-listening-in.epub with source note Listening In, and verified the Cellebrite citation of pp. 142-147 against the text: the manufacturer-relationships, reverse-engineering, jailbreak-similarity, and zero-day pricing claims all sit on the cited pages (index entry “zero-day vulnerabilities, 142-43, 147”).
  • Researched the evidentiary reliability of secret forensic methods and created Reliability of mobile forensic extraction as evidence: under Sweden’s free evaluation of evidence there is no admissibility gate, so trade-secret tools plus HD’s EncroChat non-review leave reliability practically untestable; NFC’s ISO/IEC 17025 accreditation and RåV 2022:31 certify laboratory process, not tool truth; Marlinspike’s UFED compromise makes “the tool was manipulated” a legitimate defence hypothesis; and the Bhima Koregaon Arsenal reports document malware-planted evidence with two laboratories reaching opposite conclusions. The note decomposes “found on the suspect’s phone” into content, presence, and act, and shows why self-verifying content and tool-attested generic material deserve different weight.
  • Updated Are Swedish mobile-extraction filters disclosed at trial with the NFC accreditation and RåV 2022:31 findings (partial progress on the standards task), and linked the new synthesis from Mobile-device extraction and evidentiary selection and Cellebrite.
  • Hardened the mitigations section of Reliability of mobile forensic extraction as evidence after challenge: split the controls by threat model. Chain of custody, examiner testimony, and accreditation address human tampering and cannot detect a silent software compromise (the Marlinspike manipulation was operator-invisible); the software-level controls are pre-exploitation imaging (noted as mostly impossible on mobile, where exploitation precedes imaging, so re-extracting the same device only reads back first-tool contamination), consistency-web analysis, device-external corroboration, and reproducible tooling. Also recorded the device-to-tool vs tool-to-device direction distinction.
  • Added the institutional-condition section to Reliability of mobile forensic extraction as evidence after a second challenge: every verification control is latent and activates only if someone performs it. An accredited-lab report shifts the practical burden to the least-resourced participant, and the activation conditions are worst in CSAM-class cases (visceral short-circuit, stigmatized defence task, expert-cost asymmetry, guilt narrative predating the artifact). The Bhima Koregaon scrutiny happened only under conditions almost never present, which inverts the comfort: the pipeline is most examinable where the caseload least needs it.
  • Moved note titles from a body H1 into frontmatter title, with an optional subtitle, across all 413 notes. The rendered page and Obsidian both showed the title twice, because Obsidian displays the filename above a note and the wiki emitted an H1 from the body, which the old lint rule forced to match it. The rule turned out to be enforced by lint and read by nothing: build_resolver keys on the filename stem and aliases, so wikilink resolution never saw the H1. Titles can now carry the colons, question marks, and subtitles that Windows filenames forbid, while the stem stays the link target that Obsidian and Markdown Oxide require. rumdl enforces the split without a custom rule, since it reads a frontmatter title as a top-level heading: MD041.level = 2 rejects a body H1 and MD025 rejects both. just new and just rename now take a free-form title and derive the stem, so the lossy shortening that produced SOU 2024 75 from SOU 2024:75 is no longer done by hand.
  • Added the Swedish terminal case to Reliability of mobile forensic extraction as evidence: independent defence re-examination of a CSAM-tainted image is not just unaffordable but criminal. RH 2007:21 counts deleted-but-recoverable files as possession, so the raw image is itself contraband; NJA 2014 s. 14 directs forfeiture and destruction of the media; and 23 kap. 21 a § RB lets copies be withheld for the material’s character. The only lawful examiners are the state’s own chain; the defence is left arguing over descriptions, hashes, and procedures — everything except the artifact.
  • Wrote per-type naming conventions into AGENTS.md and repaired the titles the old filename rule had deformed. A handle and a title name the same subject but need not be the same string, and each note type diverges in its own way, so the conventions are now a table rather than one rule. Agents must read the names a subject’s neighbours already use before naming a new note, which is what would have caught the drift this pass fixed. All 17 question titles carry their question mark again, and 22 notes recovered a real title or subtitle: works whose titles had been truncated at their colon, and citation identifiers such as SOU 2024:75 and C-136/17 that had been stripped of the punctuation that makes them citations.
  • Changed what happens when research settles a question. A settled question now converts to the type its answer turned out to be — concept when the answer defines something other notes argue from, synthesis when it is a conclusion the vault has a stake in, entity when it individuates a named thing — or merges into the notes that carry the finding. A question note stays a question only when the settled result is that it cannot be answered from evidence the vault can reach. Otherwise type: question means open, so just query "type=question" keeps enumerating the research frontier instead of accumulating closed history.
  • Renamed 14 notes and restored the titles on 22, finishing the repair the frontmatter migration had only relocated. Works truncated at their colon, comma splices standing in for one, em dashes flattened to a spaced hyphen, and citation identifiers stripped of their punctuation now carry the real form in title and a clean handle in the filename: SOU 2024:75 and C-136/17 read as citations again, and their stems close to SOU 2024-75 and C-136-17 rather than splitting into two numbers.
  • Recorded the case conventions, which divide into names the wiki transcribes and text it writes in its own voice. A proper noun keeps its owner’s styling, which is why secureblue, addy.io, llama.cpp, and mnemonic stay lowercase against a sentence-case filename rule that would otherwise corrupt them. A source title reproduces what the work prints, so English trade subtitles stay in title case and Swedish and academic ones stay mid-sentence. The wiki’s own subtitles are mid-sentence case, since a subtitle continues its title grammatically.
  • Corrected the source-note definition, which claimed each note describes one artifact while 66 of 129 describe several. A source note covers one work or one bundle, and bundling is preferred where splitting would produce notes that co-occur in every citation. A bundle needs one origin and one moment: artifacts sharing only a topic are a junk drawer, and the cost lands on citation.
  • Subtitles are now the exception rather than the default. Every documentation surface modeled just new with one, so an agent reading them would supply a subtitle for every note, while 404 of 419 notes have none and should not. The field carries the second half of a title that has two halves, which in practice means a source note whose work prints a colon; a synthesis handle is already a claim, so restating it as a subtitle adds nothing.
  • Fixed just new repeating the subtitle inside the title. A title carrying its own subtitle now splits at the separator, so "Securing schools, protecting minds: a scoping review" no longer stores the whole string in title beside a subtitle holding half of it again. A question mark ends a question rather than introducing anything, so a question note gets its handle trimmed without gaining a subtitle.
  • Gave titles and subtitles the smart typography the body already had. Both are emitted outside the note body, so they were escaped and never passed through smarty: a subtitle showed a literal -- and &#x27; where the body two lines below showed an em dash and a curly quote.
  • Promoted 152 parenthetical -- sequences to --- across 31 notes. smarty maps -- to an en dash and --- to an em dash, which AGENTS.md had never said, so agents writing -- for a parenthetical got the wrong mark. Closed-up ranges such as 1956--2022 are correct en dashes and all 67 are untouched, as are the -- sequences inside code fences and spans.
  • Linked Android app distribution trust models to Android Package, which defined the format five of its sentences depend on while carrying no inbound link from it. The vault types protocols and formats as concept and the products implementing them as entity, so Android Package sits with JMAP, Messaging Layer Security, AT Protocol, and ActivityPub rather than with Signal, GrapheneOS, and F-Droid. Its capitalization is the format’s own, which is why the abbreviation is APK.
  • Replaced the concept and entity boundary in AGENTS.md with an ownership test, and retyped nineteen notes from concept to entity. The old definition listed “a technical standard” as a reusable idea, which it is not: a standard is a specific document a body publishes and versions, which is the entity warranty. The clause invited agents to pattern-match a note’s prose instead of asking who owns its subject, and the vault had drifted to typing Matrix as an entity and XMPP as a concept with nothing distinguishing them. The test is now whether a specific party controls what the subject is and what it becomes.
  • Retyped the protocols and standards ActivityPub, AT Protocol, Messaging Layer Security, JMAP, XMPP, Nostr, Tox, Secure Scuttlebutt, Marmot Protocol, and Android Package; the vendor features Memory Integrity Enforcement, Apple Advanced Data Protection, and Google client-side encryption; and the networks and currencies Bitcoin, Ethereum, Monero, Zcash, Solana, and Lightning Network. This reverses the convention stated in yesterday’s Android Package entry, which recorded the drift as though it were the rule.
  • Recorded the decision in Ontology redesign proposals, which needed it: its proposed entity type vocabulary folded protocols into product because the entity inventory it was drafted from held no protocols, all of them having been mistyped as concepts. Added a standard value for published specifications, and left vendor features such as MIE under product, since Apple ships MIE and no body publishes it.
  • Left Hardware security keys and PGP and OpenPGP as concepts. The FIDO note defines a device category any vendor can build rather than the FIDO Alliance specifications, which is the general point that a body’s name in a title does not settle the type. The PGP note should become two entity notes, since its own opening separates Zimmermann’s 1991 software from the IETF format, but seven inbound links and a shared failure-modes section make that content work rather than a retype; it is queued in the proposals note.
  • Noted a gap the sweep opens: the protocol notes carried the vault’s only account of the categories they implement, so federated microblogging and memory tagging now have no note defining them.
  • Considered and rejected synthesis for Hardware security keys. Its “Catalog role” section reports judgments made in Privacy gear catalog and Swedish and EU privacy suppliers rather than asserting its own, so the note remains a building block those argue from. Flagged a mild scope smear for later: the definition of the authentication approach shares the note with commercial guidance for the shop venture.
  • Renamed FIDO security keys to Hardware security keys and rewrote its opening to name the standards it implements. The old title put a specification in the handle for a note that defines a device category: the body said “FIDO” exactly once, in the definitional sentence, and never mentioned WebAuthn, CTAP, or passkeys. The opening now redlinks FIDO2 and WebAuthn, which queues the specification pages without creating ones nothing yet calls. All seven inbound links wanted the device category and its deployment story, so none needed rewriting.
  • Corrected that rename’s handle from singular to plural within the same session. The singular followed the majority of concept handles, but the note’s own cluster runs the other way: Defence marker sprays is the same kind of subject in the same catalog, and the two pipe links the singular required in Privacy gear catalog and Privacy gear webshop were the signal that the handle fought the prose where it is used most. The body keeps its singular definition, matching Reproducible builds and Clear backpack policies, which also carry plural handles over singular openings. AGENTS.md says nothing about grammatical number, so the first choice applied an unwritten convention; the neighbor check that AGENTS.md does require would have caught it before the rename rather than after.
  • Researched privacy screens for Pixel phones and created Privacy screens for Pixel phones with a purchase recommendation grounded in four imported sources: Android Authority’s microlouver explainer, a timestamped transcript of HighTechCheck’s Pixel 10 Pro XL protector test, SkalHuset’s Spigen privacy listing, and Tullverket’s online-shopping fee rules, each with its own source note. Findings: tempered-glass privacy protectors double as screen protectors and stacking on existing glass is a mistake; brand matters mainly for louver optics, adhesive uniformity (fingerprint unlock), and install trays; PanzerGlass makes no Pixel privacy variant, leaving Spigen GLAS.tR EZ Fit Privacy as the credible pick; and since 1 July 2026 non-EU orders carry a 3 euro per-item fee plus VAT and carrier handling, so Swedish or EU dispatch is the duty-free route.
  • Updated Privacy gear catalog § Privacy screens with the Pixel supplier findings (no PanzerGlass Pixel privacy, fingerprint-sensor verification, replace-don’t-stack) for the shop’s assortment decisions.
  • Renderer gained per-family Obsidian callout icons and theme colors for all thirteen default callout families; callout markers are case-insensitive to the renderer and just lint/just fix now enforce lowercase house style.
  • Lowercased every existing callout marker and gave all 55 maintained callouts descriptive custom titles.
  • Added newly supported callout families where they fit: danger for the jammer exclusion in Privacy gear catalog, failure for seed-recovery limits in Lightning Network, quote for the eyebrow-pencil attribution in Internet Matters Online Safety Act report, example for the Cap_able/NtechLab test in Adversarial clothing, tip and success for buying criteria and drop-test results in Privacy screens for Pixel phones, todo for the telemetry re-test in Redact.dev, question for the open selection-effects questions in Swedish lay judges and political influence, and abstract for the orientation summary in Less harmful social media.
  • Converted short follow-up task lists into todo callouts in Swedish lay judges and political influence, Dumpen, Naming convicted people in Swedish media, Swedish remand detention and restrictions, Australian under-16 social media ban, EU age verification and Internet privacy, Naming the wiki, and Duress credentials and coercive extraction; longer research agendas stay as checkbox task lists. Callout titles also render slightly larger, and task checkboxes now render inside callouts.

2026-07-26

  • Imported four sources on the Mullvad donation controversy at the user’s request: Mullvad’s official statement (Mullvad donation controversy statement, saved page), Daniel Berntsson’s Örebropartiet post (Berntsson Örebropartiet post, post page plus a homepage snapshot of dberntsson.info), The Hated One’s critical video (The Hated One on the Mullvad donation, timestamped transcript), and the Mullvad segment of Privacy Guides’ This Week in Privacy #63 (Privacy Guides TWIP 63, timestamped segment transcript, first ~33 minutes only, per the user’s instruction).
  • Created entity notes for the event and its actors: Mullvad donation controversy (timeline, positions, vault implications, open threads), Örebropartiet (with an explicit caveat that the vault holds no party-primary source), and Daniel Berntsson.
  • Placed the evidence outward: Mullvad gained a donation-controversy section separating the unchanged technical assessment from the new partner-brand risk; Privacy voucher shop gained a partner reputational-diligence check; Privacy trust and governance gained owner-conduct as an explicit trust question and section; Privacy competitors and inspirations notes the market-level lesson that a VPN’s trust posture includes its owners’ public conduct.
  • Environment note for future imports: YouTube blocks the VPN exit for direct transcript API calls (youtube_transcript_api raised RequestBlocked), but the MCP YouTube tool works because mcp-youtube-transcript.exe and its helpers are excluded from the VPN via IVPN split tunneling; use that path (or those executables) for transcripts.
  • Review pass on the new controversy notes: converted the source-note “Evidentiary value and limits” sections and Örebropartiet’s “Assessment limits” into [!warning] callouts with descriptive titles, and added missing local abbreviation definitions (ÖP in Berntsson Örebropartiet post, Örebropartiet, and Daniel Berntsson; VPN and SPN in The Hated One on the Mullvad donation).
  • Imported the Guardian’s report on the Cop City phone-wipe prosecution (US government targets Cop City protester over phone operating system, rendered page plus raw companion), and created the entity note United States v. Tunick for the case itself: the first documented standalone US charge over a GrapheneOS duress-style wipe, brought under 18 U.S.C. § 2232, with a suppression ruling expected after October 2026.
  • Placed the evidence outward: GrapheneOS’s duress section gained the case and the Catalan Pixel-profiling pattern; Duress credentials and coercive extraction cites the case as the legal exposure made concrete; Encryption key disclosure compulsion adds prosecuting the destruction itself as a third US mechanism beside compelled decryption and civil contempt; Why a known harsher penalty can beat an uncertain lesser one narrows its US watchlist gap with the terrorism-watchlist placement of an uncharged protester; Ethics of resisting lawful extraction links the case where it names the evidence-destruction risk. Left Border search exception and Cop City as redlinks.
  • Followed up with TechCrunch’s next-day report (TechCrunch on Tunick duress-password prosecution, rendered page plus raw companion), which added the not-guilty plea, the attorneys’ GrapheneOS confirmation, the Defend the Atlanta Forest naming, and — most valuable — publication of the primary documents. Downloaded both filings and bundled them as United States v. Tunick case filings: the sealed November 2025 indictment (charging destruction “before and during” the search and seizure, the government’s answer to § 2232(a)’s timing question) and the March 2026 suppression motion (officer quotes with recording timestamps, the e-reader detail, the “violent extremist group” classification, and the Touset-versus-Aigbekaen circuit-split argument). Rewrote United States v. Tunick on that primary grounding, including the new mens rea gap (no public evidence Tunick knowingly entered the duress credential), and refined the § 2232 paragraph in Encryption key disclosure compulsion with the indictment’s timing language and the defense’s reliance on the 2012 Eleventh Circuit decryption case.
  • Rewrote Ontology redesign proposals from a set of open proposals into an adoption plan. Its opening still said “Nothing here is adopted” after the concept and entity boundary had shipped, and a decision was taken to adopt the remaining three in sequence. Added a status table and a four-step adoption sequence.
  • Moved decompose mode out of the distill-syntheses skill. The skill was to run it at writing time and a ratchet mode over legacy notes, sharing one procedure, but the two are different problems: an agent writing a synthesis holds the argument already and can name its premises from working memory, while recovering the premises of a note written months ago is archaeology. Write-time extraction is the demand-driven proposal itself, now shipping as an AGENTS.md default for research generally; the skill keeps only the retroactive job.
  • Rescoped that skill’s procedure from a note to a cluster. A premise seen in one synthesis looks unique to it, which is how a concept gets written to the shape of its first customer; seen across a related group it repeats in different wording, and the repetition shows both that it deserves a note and what its natural scope is. Added a step that finds repeated citations before any reading for judgment, since just clusters already partitions the vault into 30 link neighborhoods.
  • Dropped the separate hand-run pilot that gated the skill. The procedure would be run on a worked cluster while drafting anyway, so the gate named a step that happens regardless. The skill is now written by running it on the evidence-law cluster. What the gate protected against is stated directly instead: do not write the procedure abstractly and then point it at every synthesis in the vault.
  • Stated the skill’s goal as removing demonstrated duplication rather than producing a well-formed ontology. The first is checkable against a cluster inventory; the second can rationalize any extraction, which matters because extracting feels like progress and leaving alone does not.
  • Renamed the type frontmatter field to kind across the repository, the first step of the layering adopted in Ontology redesign proposals. The swap covers 431 notes, the renderer, lint, queries, the note template, the test suite, the Obsidian graph groups, AGENTS.md, the skills under .agents/skills/, and their .claude/skills/ mirrors via just sync-skills. kind now names the warranty role; type returns later as the per-kind subtype field.
  • Kept the internal JSON key as type in wiki-data.js, the search index, and the build cache. Those are a separate namespace from frontmatter, read by graph.js and wiki.js, and renaming them would have touched the browser assets for no reader-visible gain.
  • Caught one real defect the rename introduced. just new inserted a subtitle by anchoring on the literal string \ntype:, which the template no longer contains, so every new note with a subtitle would have silently dropped it. The test suite caught it; just lint did not, which is the case for anchoring on a field name in a string literal rather than reading the parsed frontmatter.
  • Added event as a ninth note kind and moved fourteen notes into it from entity: nine data breaches, the criminal cases Kevinfallet, Da Costa-fallet, and Mordet i Västermalmsgallerian, the Grindr data-sharing enforcement record, and the Mullvad donation controversy. entity had been carrying both things that persist and episodes that unfold, which is a conflation the warranty exposes: an entity note claims an identity, an event note claims an accurate account of what happened, when, and per whom.
  • Discarded two tests that had been standing in for the real one. Cultural naming does not decide it: Kevinfallet is a Swedish proper noun and still an episode. Neither does member count: a proposed enforcement subtype dissolved when its two members turned out unalike — the Grindr note narrates a regulatory action, and the Mullvad note narrates a controversy.
  • Checked the vocabulary against formal ontologies rather than settling it by feel. BFO and DOLCE divide continuants from occurrents on exactly this line, and CIDOC CRM separates intentional activity from mere occurrence, which is what distinguishes a planned festival from an unplanned controversy. The wiki keeps plain words: those standards serve machine interoperability between institutions, while this vocabulary tells an agent at a glance what a note promises.
  • Kept statutes as entities and court judgments as sources. UK Online Safety Act tracks a document that persists and gets amended, and Australian under-16 social media ban opens on an enactment but spends fifty-three of sixty-one body lines on standing obligations. The six judgment notes such as GC and Others C-136-17 each describe a saved artifact, so they stay source. AGENTS.md now states that boundary as a test: does the note describe an artifact you saved, or a thing in the world you are tracking?
  • Renamed Wiki type system from “Wiki type system” and brought it current. The 2026-07-25 field rename had left it describing a type field and eight kinds, and it still carried the swap as a future possibility.
  • Shipped the type subtype field on every entity and event note, completing the layering adopted in Ontology redesign proposals. Entities carry person (6), organization (39), product (51), standard (17), legislation (3), judgment (0), and work (1); events carry breach (9), case (3), and controversy (2). kind states the warranty a note makes; type states what the subject is, and drives targeted queries and review cadence rather than anything a reader sees.
  • Verified the query behavior before writing the vocabulary down, since the whole scheme rests on it. just query "kind=event type=breach" returns the nine breaches, kind=entity type=breach returns nothing, and lint accepts the unknown field. matches_expr ANDs every term over arbitrary frontmatter keys, so no query changes were needed.
  • Added work during the pass. Adam och övergreppen is a three-part Uppdrag granskning series, which is neither a product nor an organization, and a scan found it is currently the vault’s only entity of that shape. It is one member, and thin, but a published journalistic work had no honest home in the other six values.
  • Left judgment empty on purpose. The vault’s six court judgments are source notes describing saved artifacts, and the value exists so a ruling tracked for its effects has somewhere to go besides legislation. Neither vocabulary is lint-enforced, so an invented value fails silently; AGENTS.md says so beside the tables.
  • Removed the event subtype vocabulary shipped earlier the same day. breach, case, and controversy described what an event was about, which is the work tags already do, rather than cutting structurally the way organization and standard do for entities. The Afghan relocation data breach is the clearest evidence: an official emailed a file to the wrong recipients, so breach was already stretched past its usual sense. A topical vocabulary grows with subject matter, since a recall, an outage, or a supply-chain compromise would each want a value, while a structural one stays stable. All fourteen events are historical and equally stable, so the vocabulary earned nothing it could not get from tags.
  • Noted the pattern for future passes. This is the third category minted and withdrawn today, after enforcement and feature, and all three failed one test: their members did not behave differently. Entities keep their subtypes because those do cut structurally. The layering does not require every kind to carry a vocabulary.
  • Reverted the whole kind-and-type layering. The frontmatter field is type again across all 431 notes, the subtype field is gone from the 117 entity notes, and Wiki kind system is Wiki type system once more. Renderer, lint, template, tests, Obsidian graph groups, AGENTS.md, and the skills all moved back.
  • The layering assumed kind and type were two levels of one hierarchy: what kind of note, then what type of entity. They are not. kind asked what warranty a note makes to a reader and the subtype asked what sort of thing the subject is, which are unrelated questions in adjacent frontmatter lines. organization is not a narrower warranty than entity; it is the same warranty about a different subject.
  • Four subtype vocabularies were tried and withdrawn in one day, and all four failed the same test: their members did not behave differently. The entity vocabulary is the instructive one because it looked strongest: product held 51 of 117 notes, which is a default rather than a classification; person was recoverable from the wrongful-conviction tag; work and judgment had one member and none. Only standard and legislation changed anything, and what they changed was review cadence, which review_after already states per note more accurately.
  • Replaced the vocabulary with a table of examples in AGENTS.md. The original discomfort was real: entity covers people, agencies, protocols, statutes, vendor features, and a documentary series, and an agent meeting that breadth cold will hesitate. Listing the range with worked examples resolves the hesitation without a field to maintain.
  • AGENTS.md now records that there is no subtype field and why, so the next session does not re-derive it. Topic belongs in tags, cadence in review_after, and everything else in prose.
  • Swept the law-adjacent concept notes against the ownership test, which the 2026-07-25 pass had added without applying beyond protocols and vendor features. Eighteen notes were checked and three moved to entity: Kränkande fotografering and Olovlig avlyssning, which each open by identifying a section of the Penal Code, and Criminal-offence data under GDPR Article 10, which carries the article number in its handle and spends twenty-two of sixty body lines on the Article itself.
  • Fifteen stayed concepts, and the boundary is sharper than “enacted by a body,” since everything in a legal system traces to an enactment. The question is whether the note is about the instrument or about what the instrument established. Anonyma vittnen is the ordinary Swedish term for an arrangement introduced by prop. 2024/25:20, not the name of a provision; Resning in Swedish criminal cases describes a procedure; Free evaluation of evidence in Sweden describes a principle; Utgivningsbevis and people-search services spends most of its two hundred lines on the consequences of a certificate rather than on the certificate.
  • Recorded the instrument-versus-doctrine test in AGENTS.md beside the entity and event boundary, so the next legal note does not need the same reasoning rebuilt.
  • Shipped the demand-driven extraction default to AGENTS.md as “Extract what the task needs”, in the research workflow rather than as a maintenance skill. An agent researching a claim already holds a goal and is reading sources against it, so it is better placed to name a premise than a later pass that has to recover the argument from prose. The section carries the three drafting rules and states the scope explicitly: what the current task needs, not every idea a source touches.
  • Used Android app distribution trust models as the worked example. Its comparison rests on two premises that became Android Package and Source-to-binary correspondence instead of paragraphs restating each, and the second one shows the scope rule working — it defines the property in general rather than what one distribution channel does differently.
  • Folded the building-block shapes into that section as a prompt. These were a candidate concept subtype vocabulary earlier in the same session; they do the same nudging work as a checklist during extraction without a field to classify sixty-two notes against. The graph pass in the finishing checklist now also looks for premises the session’s own writing restated.
  • Rewrote the extraction default’s opening, which had presented its three options as a menu and then added that direct citation “is not a failure”. Read together those two passages licensed the cheapest option, which is the one requiring no thinking, and the section exists to prompt the thinking. Extraction is now the instruction and direct citation is the residue after it: a figure used once, a case that illustrates rather than establishes, a primary legal text that belongs in a source note.
  • Named the failure mode rather than only the correct behavior. A synthesis whose premises are all direct citations has usually not decomposed its argument; it has written prose with footnotes.
  • Restored the recording requirement that the original proposal called a “recorded exception” and the first draft dropped. A reusable premise cited directly is now written into the session’s Log.md entry with its reason, so a later pass can tell a considered choice from a skipped one. The balance is stated in both directions: a forced concept written to satisfy a rule is worse than a direct citation, but reaching for the citation first skips the work.

2026-07-27

  • Read the evidence-law cluster in full — Rättssäkerhet in Swedish criminal cases, False accusations as an epistemic risk, Witness reports are not ground truth, Coercive interrogation and false confessions, Why you shouldn’t talk to the police, and Mobile-device extraction and evidentiary selection — to find out what a distillation skill would actually have to do. An earlier pass in the same session worked from grep windows and frontmatter instead, and found only the premise that happened to be repeated verbatim. Full reading is the procedure; the shortcut cannot see a premise restated in different words.
  • Extracted Investigative confirmation bias as a concept. The mechanism was load-bearing in five notes across four clusters and defined in none of them, carried instead by two source cards in different domains, Moa Lidén’s Confirmation bias in criminal cases and Cognitive and human factors in digital forensics. Neither was a redlink, so just wanted never reported it.
  • Wrote the concept from the thesis rather than from the citing notes, which recovered two things every citing note had flattened. Lidén’s finding is stage-specific: police questioning, prosecutorial review, and pre-trial detention show the bias for different reasons and admit different countermeasures, so awareness training addresses at most one of the three. And the confirmatory reasoning can be rational or irrational depending on whether probabilistic or judicial rationality is applied, which makes the bias something other than carelessness. Study III is the direct warrant for separating the judge who detains from the judge who tries, a reform Rättssäkerhet in Swedish criminal cases already recommended without stating what supported it.
  • Linked the new concept from the five notes that had been restating it, including two table cells that read “Tunnel vision” as bare prose.
  • Fixed a linking failure in Android app distribution trust models, the note AGENTS.md cites as the extraction exemplar. It named F-Droid twenty-seven times while an F-Droid entity note existed, and never linked it or Reproducible builds. Extraction had succeeded and linking had not, which is a failure a “did you extract the premises” check passes.
  • Recorded both lessons in AGENTS.md: that creating a page does not connect it, and that restated premises are invisible to search because only verbatim repetition is searchable.
  • Identified but did not extract a second premise, recorded here under the rule for a reusable premise left uncited. The vault argues in at least four notes that an institution’s own outputs cannot measure its own error rate: as “finality is not ground truth” and “counting such cases measures the correction channel” in Rättssäkerhet in Swedish criminal cases, as “discovery counts are not rates” in False accusations as an epistemic risk, as the measurement framing in Witness reports are not ground truth, and as Duane’s selected sample in Why you shouldn’t talk to the police. These share almost no vocabulary. It was left unextracted pending a decision on where it belongs. A first reading suggested retyping False accusations as an epistemic risk or Witness reports are not ground truth to concept, on the ground that other syntheses reference them as building blocks. That inference was wrong. Being cited by a synthesis is not the test; a synthesis citing another synthesis’s conclusion is an ordinary and healthy edge, and the property that matters is whether the chain reaches a concept or entity at all. Sixty-three of sixty-five synthesis notes do. False accusations as an epistemic risk stays synthesis: it asserts a conclusion the vault has a stake in, and it reaches building blocks.
  • Linked Investigative confirmation bias from Witness reports are not ground truth, which was one of only two synthesis notes reaching no building block at any depth. Its sole outbound link was a source card. The amplification relationship was already asserted from the concept’s side and is now stated from both.
  • Measured where extraction is owed, and found the obvious metric misleading. Ranking syntheses by their share of concept and entity links puts Privacy venture roadmap near the bottom at one in thirteen, but that note links twelve syntheses and is doing routing work, not carrying undecomposed premises. The signal is a high count of direct source citations, not a low count of building blocks. By that measure Pornography literacy and harm reduction is the strongest candidate in the vault: eight source notes cited directly against one concept.
  • Ran just sync-skills. The 2026-07-26 revert of the kind-to-type rename never ran it, so the mirrored copies under .claude/skills still said kind: while .agents/skills said type:. The mirror is gitignored, so the drift is invisible to review and each worktree carries its own stale copy until the recipe is run.
  • Researched and imported the Google CSAM false-positive cluster at the user’s request, anchored on the NYT Mark-and-Cassio story. Seven source notes: NYT Google CSAM false positives (full-text mirror; NYT itself blocks fetch), Koffel Law Google AI flags dad (defense-practice framing), EFF Google scans false accusations (platform error-rate data: Facebook 75% non-malicious, LinkedIn 31/75 confirmed), How Google detects and reports CSAM (Google’s official pipeline, H1-2022 scale, and appeals-reform promise), Online child safety ecosystem (Stanford Internet Observatory’s 66-interview CyberTipline report, PDF), Lawshe v Verizon (Pfefferkorn on § 2258A/B immunity and the false-report lawsuit, settled and vacated February 2026), and NYT child online mistake digital life (Watkins case; raw response kept because bot protection broke the rendered snapshot). New subject notes: Automated CSAM detection (concept), NCMEC CyberTipline (entity), CSAM scanning externalizes error costs (synthesis), and the open question What is the production false-positive rate of automated CSAM detection. Linked from Less harmful social media, EU private communications law, and Alternatives to age verification; redlinked PhotoDNA, Content Safety API, and Client-side scanning.
  • Researched AzireVPN (new entity note): the Swedish WireGuard-only provider with self-owned diskless RAM-boot servers and the Blind Operator kernel module, acquired by Malwarebytes in November 2024 and third-party audited by X41 D-Sec in 2025-12–2026-01. Imported evidence: Penetration Test on AzireVPN (X41 public report PDF — no user-activity logging observed, but two critical findings and bypassable physical hardening), Malwarebytes acquires AzireVPN (acquirer press release), and AzireVPN joins forces with Malwarebytes (seller-side announcement). Placement: Mullvad gained a white-label partnerships section (Malwarebytes Privacy VPN ran on Mullvad’s network from 2019 before moving to AzireVPN infrastructure); Privacy claim assurance uses the X41 report as a worked example of commissioned verification and its limits; Swedish and EU privacy suppliers records AzireVPN as the operating-jurisdiction versus ultimate-ownership split case. Redlinked Malwarebytes and Warrant canary.
  • Imported the UK circumvention-evidence cluster: the primary DSIT-commissioned BMG survey Children’s circumvention behaviours online (2,299 UK children, May 2026 — privacy-led VPN use, 7% using VPNs for age-restricted access, false-birthdate-dominated bypass, platform substitution), plus two TechRadar reports as secondary coverage: We decided not to limit VPNs (the government U-turn and platform-duty package) and No, it’s not age verification (study detail plus industry open letter). Placement: Age-verification circumvention services gained the DSIT corroboration and the UK outcome as the first concrete test of its targeted-friction forecast (review brought forward to 2026-11-15 for the October Ofcom and ICO report); UK Online Safety Act gained a July 2026 circumvention-package section (VPN decision, 16–17 curfews and feed defaults, AI chatbot measures); VPN service legal risk cites the UK decision against a general-ban base case; Age assurance, the EPRS VPN briefing note, the Internet Matters Online Safety Act report, and Does age verification reduce harms other measures cannot each gained the corroborating DSIT evidence. Redlinked Childnet and VPN Trust Initiative.
  • Researched Scandinavian legal realism critically, prompted by the user’s gloss that the school teaches “rights do not exist” and “might makes right.” Rewrote the concept note around three documented theses (Value nihilism, rights as hollow words, law as organized force) and assessed the gloss: false as normative teaching (justification-talk is denied, not captured by power), but with real bite as practical grammar, including Berggrav’s 1945 “Hitler’s prophet” charge and Olivecrona’s wartime Schmittian use of Hägerström. Imported four sources with artifacts: The Scandinavian Value Nihilists (Strang 2009, the totalitarianism-accusation debate), Scandinavian Legal Realism and Human Rights (Strang 2018, author manuscript from Helda), Two Visions of Time (Hellner & Stenlund 2023, the rättsstat/rule-of-law paradigm shift and Hägerström’s documented impact on procedural law via Ekelöf), and An Unlikely Rights Revolution (Karlsson Schaffer, Langford & Madsen, the rights-skeptical litigation culture). Created Axel Hägerström (entity) and Value nihilism (concept), opened the question note Did legal realism weaken Swedish rights protection, and sharpened the realism paragraph in Rättssäkerhet in Swedish criminal cases: realism as documented background culture that disarmed principled resistance, not as proximal cause of evidentiary failures. Redlinked Alf Ross, Karl Olivecrona, Vilhelm Lundstedt, Ingemar Hedenius, Herbert Tingsten, Per Olof Ekelöf, and Jacob Sundberg. Bix’s “Ross and Olivecrona on Rights” and Golding’s Olivecrona appraisal were readable but not importable (Cloudflare/bepress bot walls); Nergelius 2009 on Swedish judicial review is cited from secondary literature only.
  • Cleared the six unlinked mentions just distill found on its first run. Open standard for federated e2ee email named Tuta twelve times and Proton ten while linking Delta Chat, JMAP, and Messaging Layer Security, so its two central subjects were the only ones a reader could not follow. Unmasking pseudonymous public figures now links Bitcoin, Naming convicted people in Swedish media links Dumpen, Private AI strategy links Proton, and Male variability, education, and life outcomes links Greater male variability. That last edit also emptied the report’s stranded section: the note reached no building block because its only supports were three open questions.
  • Found and fixed a counting bug while acting on the report. It scored Signal at three mentions in the federated-email note, but two were inside a footnote’s indented continuation lines, where a note title is part of a citation rather than of the argument. Only the first line of a definition was being stripped. After the fix Signal drops out correctly, since the note argues about email and cites Moxie Marlinspike’s federation claim rather than arguing from Signal.
  • Swept the vault for -- written where an em dash belongs. AGENTS.md was violating its own typography rule in three passages, each a parenthetical interruption that rendered as an en dash; the file already used --- correctly elsewhere, so the fix is what the surrounding text assumed. One note carried the same fault: HighTechCheck Pixel 10 Pro XL screen protector test quotes a video title whose saved transcript prints an em dash. Every other -- in src/ is legitimate — numeric and label ranges such as 2023--present and Studies I--III, relational en dashes such as insider--outsider, URLs, and command flags in code spans. Also merged a duplicate ## 2026-07-27 heading in this file, which a prior session appended instead of extending the dated section.
  • Imported the user-supplied Zamboni on legal realisms and politics (Zamboni 2005, ScStL 48), prompted by the user’s “might makes right” reading of its validity passage. Updated Scandinavian legal realism with a new “validity as social efficacy” analysis — the criterion is addressee-observance, not sovereign strength, and is shared with Austin, Kelsen, and Hart, but it cannot filter terror-manufactured compliance and value nihilism cannot condemn what it certifies — and upgraded the Olivecrona wartime episode with Zamboni’s footnote 61: Olivecrona publicly supported the full validity of the Nazi regime as a legal order in England oder Deutschland? (Lübeck, 1941). A follow-up exchange on Lundstedt’s common-sense-of-justice advice added the analysis that the framework is structurally majoritarian only as prudence (efficacy anchored in prevailing sentiment), leaving minority protection without a conceptual home. The Zamboni source note uses the short handle Zamboni on legal realisms and politics with the work’s full title in frontmatter.
  • A third exchange on the circularity of the validity criterion and the strawman charge against rights-talk added a section to Scandinavian legal realism: the framework lives on inherited rights-culture capital it declares meaningless; its referential semantics misses the practice/institutional-fact account of rights (Golding’s monetary-unit point, Hart’s rule-following reply via Bix); and the constructivist middle — rights as deliberately built anti-authoritarian devices — is compatible with everything true in the realist ontology.
  • Ran the distillation pass from the distill-syntheses skill. Scope: every note in the just distill report plus the correction-channel premise hunt across clusters 3, 8, 16, and 27. Eleven report-flagged notes were read in full by delegated readers; clusters 1, 2, 4, 5, 6, 9, 10, 12–15, 17–19, and 21–33 were not read end to end.
  • Extracted Correction channel as a concept. This is the premise the earlier session today left pending: the vault argued in at least six notes across four clusters that an institution’s own outputs cannot measure its own error rate, as “finality is not ground truth”, “counting such cases measures the correction channel”, “discovery counts are not rates”, Duane’s channel-selected sample, and the calibration callout in Probabilistic interpretations of beyond reasonable doubt. The term “correction channel” was already in use in six notes and defined in none. The concept is written from the sources (Lidén’s archive study is the direct warrant for rejecting exonerations as a denominator; the two JK commissions from the case-file side; Gross et al. for exoneration-sample selection) and is now linked from ten notes, including the person notes Billy Butt, Esa Teittinen, Joy Rahman, and Samir Sabri.
  • Extracted Insider-outsider model (concept), restated in bare prose in Boomers and democracy, Swedish Election Study 2006, Dualization of social rights in Sweden and beyond, and Welfare-state dualization; all four now link it.
  • Extracted WireGuard (entity), named as bare prose in four notes while being argued from as the mature-protocol recommendation; and Lagen om elektronisk kommunikation (entity), the statute three notes cited as “LEK” with no page.
  • Extracted Privacy screen protectors (concept). The general mechanism — microlouver optics, two-way vs four-way, glass vs film, the no-stacking rule, the residual-threat boundary — was trapped inside Privacy screens for Pixel phones, a page named for the idea’s first customer. The Pixel note now keeps the tested recommendation, the Pixel fingerprint-sensor specifics, and the Swedish import-fee analysis, and points at the concept for the general material; Privacy gear catalog links it too.
  • Connected pages that existed but were never pointed at, the failure the skill names with the F-Droid example: Age-period-cohort identification from Boomers and democracy; Swedish encryption and passcode disclosure law, Swedish right to silence and förklaringsbörda, and Coercive interrogation and false confessions from Swedish remand detention and restrictions; Utgivningsbevis and people-search services and Data minimization from EU and Swedish data removal rights; Data minimization from Age assurance; Criminal-offence data under GDPR Article 10, IMY right to remove search results, Offentlighetsprincipen and personal privacy, Protected personal data in Sweden, Utgivningsbevis and people-search services, Dumpen, and Integritetsskyddsmyndigheten from Challenging name-search indexing of criminal coverage; Clear backpack policies from Clear backpacks are school security theater; four rättssäkerhet sub-mechanism pages from Scandinavian legal realism; Privacy threat modeling and Ofcom from VPN service legal risk; Greater male variability from Male variability, education, and life outcomes, whose restated evidence section was compressed into the link; Dumpen from Naming convicted people in Swedish media; Proton, Tuta, and Signal from Open standard for federated e2ee email; Bitcoin from Unmasking pseudonymous public figures; Proton from Private AI strategy.
  • Left as direct citation or redlink, with reasons: security theater is a single-cluster borrowing so far, redlinked as Security theater rather than extracted; the median-voter model has only two thin in-vault groundings, redlinked as Median voter; the US-doctrine cluster around Miranda, Marks, and Davis in Invoking the Fifth Amendment can itself be used as evidence is a real extraction debt but a larger job than this pass, queued as Miranda warning; a GDPR entity note and an 18 U.S.C. § 2258A statute note are genuine gaps, likewise deferred. The report’s remaining entries (Pornography literacy and harm reduction, Naming convicted people in Swedish media, Age-verification circumvention services, Greater male variability, Automated CSAM detection, CSAM scanning externalizes error costs) were read and judged correct as they stand: literature syntheses and legal-framework notes whose supports genuinely are source cards.
  • Wanted-page triage emptied the just wanted queue (28 targets) by creating 27 notes and repairing 2 links. New entity notes: Alf Ross, Karl Olivecrona, Vilhelm Lundstedt, Herbert Tingsten, Ingemar Hedenius, Per Olof Ekelöf, and Jacob Sundberg (the Scandinavian-realism figures, grounded in the saved Strang articles plus SBL, NE, and lex.dk); Cop City (working, review_after 2026-12-01; documents the RICO case’s December 2025 dismissal and the pending appeal); Comprehensive School Threat Assessment Guidelines and K-12 School Shooting Database (school-safety infrastructure); FIDO2, WebAuthn, PhotoDNA, Content Safety API, Childnet, Malwarebytes, VPN Trust Initiative, Whonix, and PTS. New concept notes: Border search exception, Miranda warning, Security theater, Median voter, Client-side scanning, and Warrant canary. New event notes: Columbine High School massacre and Marjory Stoneman Douglas High School shooting. Repaired links: the Log’s stale Internet Matters Online Safety Act report reference (was “…report 2026”, a pre-rename title), and the lowercase Border search exception redlink in United States v. Tunick and the Log, which was a casing slip rather than a different subject. The graph improved because every link in the queue now resolves to a grounded note: the realism cluster’s seven named figures are individuated instead of floating as bare names, the school-security cluster’s anchor events and models exist, the CSAM and authentication clusters separate owned standards from reusable mechanisms, and the US criminal-procedure cluster gained the two doctrines its case notes argued from. Graph pass added first-mention links from The Pinch, Patch clear backpacks debate overview, secureblue features and FAQ, six Fifth-Amendment-adjacent notes to Miranda warning, Desktop operating system security comparison, Kicksecure, and Qubes OS to Whonix, and the CSTAG model name in both school-safety notes. Also corrected Children’s circumvention behaviours online: the Childnet VPN study was fielded in November 2025 but published 1 December 2025. No question notes were needed — every target grounded — and no redlinks remain.
  • Added the missing abbr definitions to secureblue features and FAQ: DNS, DNSSEC, FAQ, FIDO2, LUKS, PIN, TLS, TPM2, and UX alongside the existing OCI and SELinux. GNOME, KDE, USBGuard, OpenPAX, and sysctl were judged names or identifiers rather than abbreviations and left undefined.
  • Ran a distillation pass over the child-safety and age-assurance cluster and extracted two premises the vault had been restating without a page. Automation bias was load-bearing in four notes in three vocabularies. Automated CSAM detection wrote “inherits the classifier’s framing”, CSAM scanning externalizes error costs wrote “review confirms rather than corrects”, Content Safety API wrote “inheriting the classifier’s framing”, and Client-side scanning had only a bare “second opinion”. The vault had no page for any of it, so every note re-derived the mechanism from its own cases. Written from Goddard, Roudsari and Wyatt’s systematic review of 74 studies and Ben Green’s survey of 41 human-oversight policies, not from the citing notes. Reading Green recovered two things the citing sentences had flattened: discretion fails in both directions, since judges override risk scores toward detention and more often for Black defendants with identical scores, so a human step redistributes error rather than removing it; and an oversight requirement can legitimate the system it supervises while shifting accountability onto the reviewer, which is the general form of the cost allocation CSAM scanning externalizes error costs had only argued for CSAM. Linked it from Facial recognition, and in both directions with Investigative confirmation bias, which had reached the same conclusion about vigilance as a control from an unconnected literature.
  • Displacement was the second, restated across seven notes as “displacement to less accountable services”, “migration to unregulated or fringe services”, “the substitution margin”, and “movement to riskier services”. Written from Guerette’s COPS guide, now preserved as Analyzing Crime Displacement and Diffusion, which supplied what the vault’s usage lacked: the five forms, the benign and malign distinction, diffusion of benefits as the complement, and the empirical record — displacement in 26% of 574 observations across 102 evaluations, diffusion in 27%, and relocation usually smaller than the treatment effect. That record corrects the vault’s own habit of naming displacement as a presumed cost of age gates. Less harmful social media already had the right instinct (“displacement must be measured rather than assumed away”), while Alternatives to age verification and Pornography literacy and harm reduction were asserting it. Both now carry the base rate and the measurement duty. Children’s circumvention behaviours online holds the only measurement, and it is target and offence displacement rather than tactical, which is why enforcement aimed at VPNs addresses the smaller margin — now stated in Age-verification circumvention services.
  • Opened Does online displacement behave like crime displacement because the extraction exposed a gap rather than closing one: the base rates were measured on offenders and physical places, and the vault applies the concept to ordinary users whose substitute is one tap away. Search cost is the parameter the criminological finding depends on, so the numbers may not transfer in either direction.
  • Left three premises as direct citations after considering extraction. The Marks v. United States narrowest-grounds rule in Invoking the Fifth Amendment can itself be used as evidence is genuinely single-use here. Account-concentration risk, named in CSAM scanning externalizes error costs, is reusable and probably wants a page, but grounding it needs the privacy-continuity cluster rather than this one. The variance-ratio and tail-ratio relationship in Greater male variability is a real statistical premise that the note states adequately in place; a separate page would have one customer.
  • just distill fell from six entries to four. Naming convicted people in Swedish media, Greater male variability, and Invoking the Fifth Amendment can itself be used as evidence remain listed and are the report’s characteristic false positive: notes whose supports are properly primary law or primary studies. Scope read: cluster 3 end to end plus the out-of-cluster report entries. Clusters 1, 2, and 4 through 27 were not read this pass.
  • A fourth exchange on the realism thesis added the structural-homology formulation to Scandinavian legal realism. The user argued that Swedish mechanisms — exclusion-free evidence law, party-nominated lay judges, and a reopening structure whose failures only a powerless minority feels — are not merely tolerated by the doctrine but are its institutional shape. The note now states the middle position between causation and epiphenomenon: the mechanisms do not descend from the doctrine (nämndemän are medieval), but validity-anchored-in-prevailing-sentiment predicts their shape, and value nihilism removed the vocabulary in which they could be objected to as rights violations. Free evaluation of evidence in Sweden gained the rights-anchored contrast (the American exclusionary rule attaches the right to an evidentiary consequence; the Swedish two-track design gives a process violation no evidentiary price), and Resning in Swedish criminal cases gained the minority-constituency analysis: a majority-sentiment system has no internal channel that registers a wrong the majority never feels, which is why corrections arrive through journalism. Did legal realism weaken Swedish rights protection now records homology-plus-disarmament as a third causal shape between its two horns.
  • A fifth exchange sharpened the lay-judges point from rationalization to conservation. The user argued that without realism Sweden might have abolished nämndemän long ago; verification found the record compatible: jury proposals failed pre-realism (1832, 1849), every depoliticization attempt since 1945 has failed (SOU 2013:49’s free quota dropped from prop. 2013/14:169; the March 2024 ministerial answer called nomination reform “not aktuellt”; MP motion 2024/25:1309 is pending), the current attack runs in rättssäkerhet vocabulary and the defense in the democratic-representational register realism scientized, and Norway abolished its jury in 2017 on an explicit rule-of-law rationale, re-keying lay participation to random selection. Swedish lay judges and political influence gained a “reform history and survival” section with two imported riksdagen artifacts (motion 2024/25:1309, answer 2023/24:703), and Scandinavian legal realism now states the conservation hypothesis with its competing keystone (party patronage, which needs no jurisprudence) honestly attached.
  • Ran a whole-vault structure audit against just links and just clusters. The report’s “unreferenced” section emptied and dead ends fell from eight to six, the remainder being source cards absorbing links by design. Six source and entity notes stated their own relevance while the notes they informed never pointed back: United States v. Tunick listed both its reporting URLs in frontmatter and linked neither US government targets Cop City protester over phone operating system nor TechCrunch on Tunick duress-password prosecution; secureblue cited the two project pages its whole security-model section paraphrases without linking secureblue features and FAQ; and the same shape held for UK Online Safety Act and Fortune Gen Alpha age-verification report, CSAM scanning externalizes error costs and Koffel Law Google AI flags dad, Clear backpacks are school security theater and K-12 School Shooting Database, and Male variability, education, and life outcomes and War on Boys.
  • A mechanical check of the same defect — a note citing a URL in sources that a maintained source note owns, without linking that source note — reports 67 further instances, which is the audit’s largest open finding. Each needs reading rather than a sweep, since a direct citation is correct where the source note is a multi-artifact bundle and a link would point a reader at documents that do not support the claim.
  • Four URLs were claimed by two or three source notes at once, which made the vault’s provenance map ambiguous about which card records which artifact. Virtual private networks and child protection EPRS carried two URLs belonging to Internet Matters Online Safety Act report and Children’s circumvention behaviours online, the Internet Matters card carried the DSIT study’s URL, and GC and Others C-136-17 carried Conseil d’État 401258’s decision URL. All four now sit only in the card that describes the artifact; the prose links that already did the work correctly were left alone.
  • United States v. Tunick was retyped from entity to event. AGENTS.md names a criminal case as an event, the vault already types Kevinfallet, Da Costa-fallet, and Grindr data-sharing enforcement that way, and the note narrates a stop, an indictment, a plea, and a pending ruling rather than describing something wholly present at any moment.
  • Investigative confirmation bias argued the Correction channel premise in its own words — the correction route running through the institutions whose work is in question — without linking the note that holds it. It now does.
  • The three question notes promoted from Male variability, education, and life outcomes on 2026-07-24 were seed and, in two cases, linked nothing at all, so the vault’s research frontier had entries a reader following links could not reach. All three now link their evidence and are working.
  • Talkie and Introducing talkie are a two-note island that no subject note reaches, an import whose placement pass produced no connection. The vault holds nothing on evaluation methodology to connect it to, so the note now redlinks Temporal leakage, the named mechanism that would be its first real neighbour.
  • Audited Index as a navigation surface by computing what a reader can actually reach by following links from it. Excluding Log, which is a change record rather than a route, 30 subject notes were unreachable, and they formed two coherent domains rather than scattered stragglers: the school-security cluster around Clear backpacks are school security theater (16 notes), and the boys’-education cluster around Male variability, education, and life outcomes (10 notes). Both passed the entry-point test — durable, own reading path, distinct starting intention, and not reached through any listed hub — so both were added to Topics, taking it from seven entries to nine. Reachability from the Index is now 496 of 498.
  • Neither domain needed a new type: index page. Each already had a synthesis routing nearly its whole cluster, and an index-shaped layer over a working reading path would have duplicated it. The two remaining unreachable notes are Talkie and Introducing talkie, a two-note island too narrow for a top-level entry; they stay discoverable through Tags, Graphs, and search, which is where a cluster this small belongs.
  • Swedish business tax map was reconsidered and retained. It is reachable from Privacy venture atlas, which normally argues against a separate entry, but its content is general Swedish business taxation with nothing privacy-specific in it, so a reader arriving with a tax question should not have to enter through a venture hub.
  • Renamed the Graphs story maps from “island” to “cluster” and added four new ones, taking the authored layer from three maps to seven. “Island” described a region cut off from the rest of the graph, which is what a bridge-poor neighborhood looks like from the outside but not what these maps are: each selects a reading path through a densely connected region, and the term now matches the vocabulary just clusters already uses. The renamed heading is linked from an earlier entry, so that reference was repointed to Graphs § Business cluster.
  • The four new maps are the rättssäkerhet, child safety, Swedish public-record, and school security clusters. They cover the domains that carried the most notes and had no authored map at all: the rättssäkerhet map runs from evidence doctrine through the three mechanisms that corrupt investigative inputs to Correction channel, where the domain’s central claim sits; the child safety map hangs the design, literacy, circumvention, and scanning branches off Alternatives to age verification, which is where the domain’s comparison actually lives; the public-record map runs from Offentlighetsprincipen and personal privacy to a commercial product and back out through the removal routes; and the school security map shows an event supplying the salient case while K-12 School Shooting Database supplies the denominator that makes it rare.
  • Every edge in all seven maps is now a wikilink that exists in the notes. That was not previously true: the privacy cluster drew seven edges with no link behind them in either direction, so a reader following the map by clicking would reach a node with no route onward. It was rewired to real edges without changing the story it tells.
  • The ontology diagram was missing event entirely, although the renderer has styled it, legended it, and given it a --graph-event colour for both themes all along. The gap surfaced because Marjory Stoneman Douglas High School shooting is the first event node any authored map has needed.
  • The curated overview was stale by two entries and now carries all nine Index topics.
  • Ran a whole-vocabulary tag pass. 293 tags became 282, the near-duplicate-spellings section emptied, and the used-once queue fell from 104 entries to 92.
  • Six merges, only one of which any report section named. standard into standards was the lexical near-duplicate the tool found. The other five were spelling-divergent synonyms applied to notes that never overlapped, so no co-occurrence view could surface them: cybersecurity into security, audit into security-audit, mobile into mobile-security, hardware-security into security-keys, and regulation into law, which gives Ofcom the treatment PTS already had.
  • Two tags held two unrelated meanings each. isolation was a clean 4/4 split between virtual-machine isolation on Qubes OS, Whonix, and Desktop operating system security comparison, and the isolation of Swedish remand detainees; a reader landing on that tag page got both. The detention notes now carry solitary-confinement, and isolation keeps the sense the security literature uses. inference held only model inference while causal-inference sat beside it in the same vocabulary, and a privacy vault also uses the word for inference attacks, so it became model-inference.
  • Six singletons retired for restating their own note: power, customers, validation, migration, and workplace each repeated a word from the title of the one note carrying them, and society named a subject too broad to separate anything.
  • wiki looked like the same case and is not. The report counts it as used once because tag statistics exclude meta notes, and eight of its nine notes are meta. It is the only thing grouping the wiki’s own documentation, so the deletion was reverted. A count from a filtered view is not a count of the vault.
  • Three tagging omissions closed. Proton and Proton Lumo security model gained the Proton tag, which had been sitting on Lumo alone while the vendor page listed one of its three notes. Pegasus spyware gained security and mobile-security, which every one of its siblings already carried, and Propensity evidence from legal but suggestive material gained criminal-justice, since it describes criminal evidentiary regimes for sexual offences.
  • Distinctions deliberately preserved: evidence, evidence-law, and evidence-synthesis are epistemics, doctrine, and research method; procedure and criminal-procedure differ because Free evaluation of evidence in Sweden is a civil-and-criminal principle; and adult-content names the regulated category that age gates gate, which is not the medium pornography studies.
  • Ran a lifecycle and freshness pass. Nothing was due and nothing was stale, so the freshness half had no revalidation to do; the next scheduled review is GrapheneOS Pixel procurement on 2026-08-13. The work was in the lifecycle half: seed 2, working 134, stable 362 became seed 0, working 133, stable 365.
  • Should innocent people talk to the police carried a real material gap and it turned out to be closeable. The 2026-07-22 import fell back to an untimestamped transcript because YouTube IP-blocked the timestamped fetch, which left every claim-level citation in the note unable to point at a moment in the video — the thing AGENTS.md requires a timestamped transcript for. The fetch succeeds now, so the video’s timestamped transcript was preserved as a new artifact and the note now cites it. The original untimestamped capture stays beside it, since src/sources/ is immutable and the earlier file is what the 2026-07-22 session actually saw. The new transcript is also complete, where the old one omitted the sponsor read and the presenter’s account of his sister’s murder.
  • Two source cards were working for a reason that belongs to the freshness system rather than the lifecycle one. More flexible remand and prison enforcement describes SOU 2026:30 accurately and completely; what is pending is the legislative process, which its review_after: 2027-01-15 already tracks. Riksrevision audit of Swedish lay judges states its own scope — the audit’s remit and the current institutional design, not its eventual findings — and is reliable at that scope, with the October 2026 report tracked by review_after: 2026-11-15. Both are now stable. A due review creates a verification task; it does not make a note unreliable in advance.
  • The two remaining seeds were promoted to working. Does online displacement behave like crime displacement states its question, its two competing mechanisms, what is known, and what would settle it, and it links four notes; Naming the wiki has candidate sets, frontrunners, and an open decision expressed as tasks. Neither is a skeleton, and seed was describing their age.
  • Remaining lifecycle work, not guessed at: 63 of the 133 working notes state no gap in a task, a callout, or an open-questions section. That is a reading queue, not a verdict — a gap can be stated in ordinary prose — but it is where the next pass should look, and it is too large to resolve by inspection in one sitting.
  • Worked the lifecycle backlog the previous entry deferred. The question behind it was whether those notes would stay working forever, and the answer split: 33 of the 63 carried a review_after that will fire and force a reassessment, while 30 had neither a review date nor a stated gap and nothing on the vault’s current mechanisms would ever ask again. Three of those 30 were question notes, where working correctly means open and just query "type=question" is already the queue. The remaining 27 were read.
  • Twenty-one were promoted to stable. Eight were compact cards describing fixed technical subjects — JMAP, Ethereum, Solana, Zcash, Android Package, PGP and OpenPGP, WireGuard, and Reproducible builds — and four more were durable concepts and one statute: Insider-outsider model, Source-to-binary correspondence, Confluence model of sexual aggression, and Lagen om elektronisk kommunikation. Three were closed wrongful-conviction cases whose accounts run through acquittal: Esa Teittinen, Joy Rahman, and Samir Sabri. Greater male variability states the X-chromosome mechanism’s lack of experimental demonstration as a limitation of the hypothesis, which is a limitation represented rather than a gap in the note.
  • Five of those promotions also gained a review_after, because their subjects move even though the notes are reliable now: Eustella, F-Droid, and EncroChat on 2027-01-28, Ofcom on 2026-11-30 as the Online Safety Act duties roll out, and Pegasus spyware on 2027-01-28. Status and freshness are separate systems, so a note can be reliable today and still need a date.
  • Six venture notes stay working, and this is the answer to the question rather than an evasion. Lawful digital-safety support, Post-seizure digital recovery, Private AI services, Privacy business thesis, Privacy business proposals, and Privacy customer segments rest on commercial assumptions no calendar resolves: who buys, at what price, and whether the job is real. working is the accurate label until the venture tests them, not a defect to be cleared. Privacy business proposals gained a 2027-01-28 review anyway, since its stated SEK prices are market facts that date independently of whether the hypotheses are validated.
  • The result: working 133 to 112, stable 365 to 386, and the set with no forcing function at all fell from 30 to 8 — three open questions and five unvalidated business plans, each of which is where it should be.
  • Researched How does Swedish law treat duress wipes, the question with the most backlinks in the cluster this session had already been working. Three of its four branches now have answers, and the pass corrected three errors in what the note previously held.
  • The note had analysed the wrong statute. It examined skyddande av brottsling (BrB 17:11), found the “another person’s crime” requirement, and concluded that Swedish law probably reaches a duress wipe through no provision at all. It never considered överträdelse av myndighets bud (BrB 17:13), which punishes anyone who “olovligen rubbar, skadar eller annars förfogar över egendom som är föremål för … beslag” with fines or up to one year and asks nothing about whose crime is concealed. That is the Swedish analogue of the 18 U.S.C. § 2232 charge in United States v. Tunick, and the comparison shows a gap rather than a match: the American statute reaches destruction undertaken to prevent a seizure, while 17:13 attaches only once the property is seized. Timing, not the wipe, decides the criminal exposure.
  • The note’s reading of NJA 1990 s. 175 was backwards. It treated the acquittal of a man who helped his brother dispose of a knife as evidence that 17:11 fails even for physical evidence destruction. The case turned on the straffrihetsregel for close relatives, now 17:11’s fifth paragraph. An exemption presupposes the conduct is otherwise caught, so the case confirms the provision’s reach. The note also gave outdated penalties; Lag (2022:792) raised 17:11 to two years, and six for grovt.
  • The self-incrimination inference was also wrong, and correcting it changed the answer rather than a detail. The note reasoned that criminalizing destruction would sit in tension with the passcode privilege. Swedish and Convention doctrine already draw the line that assumes away: the privilege protects passivity, not a right to obstruct, and Saunders excludes material existing independently of the accused’s will. Refusing a passcode withholds help; wiping destroys evidence. Swedish encryption and passcode disclosure law now carries that boundary explicitly.
  • The most consequential branch was the one the note had barely developed. RB 24:5 a § permits restrictions on a detainee’s contact “endast om det finns risk för att den misstänkte undanröjer bevis eller på annat sätt försvårar sakens utredning,” and kollusionsfara in RB 24:1 is worded identically. A demonstrated wipe is not a fact from which that risk is inferred; it is a completed instance of the conduct the test names. Given what Swedish remand detention and restrictions records about Swedish isolation practice, that is a heavier consequence than the one-year maximum.
  • Imported Förklaringsbördan i brottmål, Nordlander’s 2017 Svensk Juristtidning article. Swedish right to silence and förklaringsbörda had been relying on it as “a 2017 law-review treatment” in bare prose for six days, and now links the card. Its scope also settles the adverse-inference branch: the Murray safeguards constrain inferences from silence, so a wipe, being conduct, is weighed under ordinary free evaluation without them — easier to use against a defendant than silence, which inverts the intuition that the two are one strategy.
  • The question stays working. What decides real cases is untested: whether destroying data counts as disturbing the seized property, and when a phone becomes “föremål för beslag” during a border or arrest search. Both are now tasks in the note, with Bäcklund’s commentary still unread.
  • Researched Will Swedish people-search services survive GDPR enforcement, the second question of this pass. It cannot be closed, but two of its five open tasks now have answers, and the answers point the opposite way from the doctrine.
  • The judgment landed on 9 July 2026 and changed nothing yet. Nineteen days later, Attunda tingsrätt has not applied C-199/24 to Lexbase, IMY’s four supervisions are all still open, and IMY has been declining to investigate individual complaints against those same four services since March 2026, closing them without action subject to appeal. A person seeking removal from Lexbase, krimfup, Upplysning.se, or Mrkoll today has a worse administrative route than before the ruling, because the regulator centralised the question and paused the queue. IMY suspension of search-service complaints preserves that position, and EU and Swedish data removal rights now names the three routes that still function.
  • The note’s legislative claim was out of date. It held that the government had signalled an ordinary-law inquiry but that none had been launched. Konstitutionsutskottets betänkande 2025/26:KU18 records that the replacement inquiry is running and reports by March 2027, covering background checks and search-service regulation. That sets a floor: no legislative resolution before 2027, with a proposition and passage after it.
  • The four IMY supervisions are not one question. Lexbase and krimfup are examined for the publication itself; Upplysning.se and Mrkoll for refusing erasure on request. Only the first line decides whether the services can exist, so a clean win on the second would leave the databases standing. The note now separates them.
  • The finding widens rather than narrows the venture niche, which is why it went into Privacy business proposals. The utgivningsbevis triage line exists only while removal is hard, and removal is currently harder, not easier, than before the judgment. It narrows again if IMY’s review ends in deletion orders a customer can cite without help.
  • The question stays working with a 2026-12-15 review. What would close it is the Attunda judgment and the first concluded IMY supervision, neither of which is in the vault’s control.
  • Continued the question pass with a triage of the whole frontier, then closed the two questions whose answer is that the evidence is absent. AGENTS.md reserves stable on a question note for exactly that result, and it is the one outcome no other note type carries.
  • What is the production false-positive rate of automated CSAM detection is closed as structurally unanswerable, and the reason is better than “platforms decline to publish.” The EU compelled the disclosure: DSA Article 15(1)(e) requires “indicators of the accuracy and the possible rate of error” of automated moderation. It has not produced the number for two independent reasons. The metric the Act names is uninformative at CSAM prevalence, since under that class imbalance a detector that flagged nothing would score near-perfect accuracy; Operationalizing content moderation accuracy in the DSA was imported for the argument and its worked case, where 95 percent accuracy sits beside 58 percent precision and 33 percent recall. And reporting under the provision is documented as “essentially meaningless,” with the Transparency Database no help either, since over 99.8 percent of statements of reasons cite terms of service and the free-text field is boilerplate in 98 to 99 percent of submissions. The figure exists internally and is withheld, which a 2019 expert panel commissioned by Meta already asked for and did not get.
  • That finding went into Automated CSAM detection as a reading rule, because it disciplines every accuracy claim made for this technology, including the Commission’s “significantly above 90 percent” in the CSA regulation debate. An accuracy percentage offered for a rare-target detector says less than its size suggests.
  • What interventions reduce school weapon carriage is closed on a negative answer: none, on comparison-group evidence. The note already contained that finding from a 2025 scoping review; what this pass added was an independent scholarly search of 2015–2026 that failed to contradict it, and the shape of what the search did return — school-to-prison pipeline, surveillance, discipline, and perceived safety, all easier to measure than the outcome the policies claim. The nearest rigorous design, Eisman et al.’s 2020 cluster randomized trial, is real but runs in elementary schools and measures bullying and mental health, so it shows the design is feasible while its outcome set shows why the gap persists.
  • The remaining sixteen questions were triaged rather than researched. Eight are answerable from published literature and were left for later passes; six are blocked on events or on material the vault cannot reach, including Swedish preliminary-investigation files and pre-and-post data on bans that took effect this month; and two, on legal realism and on cohort-weighted policy, need work that is closer to historiography and data analysis than to search.
  • Worked the gender-and-education question cluster, three questions sharing one literature.
  • ADHD relative-age effect is answered and is no longer a question. It asked whether the relative-age gradient in ADHD diagnosis reflects overdiagnosis of the young or underdiagnosis of the old. Two lines of evidence separate them. Kazda and colleagues’ scoping review of 334 studies applies a formal overdiagnosis framework and concludes that ADHD is overdiagnosed, with the excess concentrated among mild presentations — the signature of overdiagnosis rather than better case-finding. And Root and colleagues, following 1,042,106 UK children, find the same gradient for intellectual disability (1.30) and depression (1.31) as for ADHD (1.36), which no ADHD-specific underdiagnosis story can produce. The note was renamed and retyped from question to concept, the first such conversion in the vault, and Overdiagnosis of ADHD in children and adolescents was imported.
  • The binary the question posed was itself incomplete, which is the more interesting finding. Being youngest in a year is associated with worse grades and worse peer relationships, so part of the gradient may be difficulty that is real rather than immaturity that is misread. On that reading the youngest are both overdiagnosed and genuinely worse off, and age-norming the criteria would fix the label while leaving the disadvantage in place.
  • Is the gender-equality paradox a measurement artifact is substantially answered, and the answer is yes. The STEM version rests on a quantity Stoet and Geary had to issue a corrigendum to describe correctly: not the share of women among STEM graduates, as the paper said, but a constructed propensity ratio. More seriously, Deconstructing the gender-equality paradox decomposed the difference-score correlation the whole literature relies on and found the headline results driven by one component, men’s and women’s country means correlating at .93 to .97. When two series track that closely, their difference is dominated by variance rather than divergence, and that residual is what has been read as a phenomenon.
  • That required correcting Male variability, education, and life outcomes, which had used the paradox to argue that the pure social-construction account is harder to sustain. It cannot currently carry that weight. The synthesis also cites Herlitz and colleagues (2024) as support, while the deconstruction paper cites the same review among works raising the difference-score alarm, so the vault’s summary of it is probably one-sided. Reading Herlitz directly is now a task on the question note. The honest position is that the synthesis has one fewer limb, not that the opposite conclusion holds.
  • Does school behaviour grading cause later labour-force exit in men stays open, with a sharper account of why. The literature the question would draw on measures the trait and the outcome; the question asks about the grade, which is a mediation question needing a design where the grade varies independently of the behaviour producing it. Grading-policy changes, teacher variation, and relative-age variation are where to look, and none has been applied to this outcome.
  • Researched Does pornography use cause adolescent sexual aggression, the question the vault had been answering secondhand from two reviews that each described the dispute without holding either side. Both meta-analyses are now imported in primary form: Wright Tokunaga and Kraus pornography aggression meta-analysis for the affirmative case, Ferguson and Hartley pornography aggression meta-analysis for the sceptical one.
  • They disagree less than the secondhand accounts implied, and the disagreement is mostly about what “pornography” names. Wright and colleagues found consumption associated with aggression across 22 studies in seven countries. Ferguson and Hartley, stratifying by study quality, found no association for nonviolent pornography, a weak one for violent pornography that cannot be separated from selection, and particularly weak longitudinal evidence. The defensible position is that pornography in general is not supported as a cause, that violent material carries a weak association of unknown direction, and that the longitudinal design causality would require is the thinnest part of the base.
  • Two properties of the literature limit what any of it can carry. The meta-analytic evidence is largely adult general-population work, so applying it to adolescents is an inference rather than a finding, and adolescent-specific longitudinal research is exactly what is missing. And effect sizes shrink as method improves, which is the signature of citation bias and means a single strong positive result here deserves less weight than its statistics suggest.
  • The answer does not do what the question was posed to test. It was raised because a causal finding would strengthen the case for content regulation beyond what the vault’s intervention notes grant. It does not strengthen it. What it supports is narrower: attention to violent genres specifically, and to individuals in whom risk factors converge, rather than restriction justified by pornography as such. Pornography literacy and harm reduction now records that the violent-versus-nonviolent distinction is the load-bearing one in its list of categories, more so than the minor-versus-adult distinction most policy argument turns on.
  • Researched Does pornography literacy improve adolescent outcomes, the question the vault raised against itself. It had noted that its own standard — outcome evidence rather than intuitive confidence — binds its preferred intervention too, and that the literacy case rested on blocking’s weakness rather than on measured effects. That self-assessment turns out to be half wrong.
  • The pornography-specific evaluation base is as thin as assumed, but the method’s parent literature is not, and the note had been looking in the wrong place. Media literacy interventions meta-analytic review pools 160 interventions across forty years and finds media literacy works on average. The effects sort in a way that matters here: largest for media-relevant outcomes, smaller for behaviour, strongest for knowledge, and while most outcomes hold at follow-up, behaviour decays. A harm-reduction claim must travel from knowledge to behaviour, and that is the stretch the evidence covers worst.
  • Two further findings closed the comparison the note wanted. Aznar-Martinez and colleagues reviewed sexuality-education programmes whose effectiveness had been assessed and found pornography a neglected topic across all 30, so the literacy and education branches cannot be compared on pornography outcomes; the education branch mostly does not measure them. Scull and colleagues’ Media Aware trial is a genuine controlled evaluation but reports immediate effects only.
  • Pornography literacy and harm reduction now carries the correction. Literacy is well evidenced for changing how young people understand what they see and poorly evidenced for changing conduct, which is narrower than “harm reduction” implies. It is still more than blocking can show, which has neither kind of evidence, but the asymmetry the note assumed is smaller than it claimed.
  • Researched Which social-media designs reduce harm for young users, the design branch that Less harmful social media and Alternatives to age verification both stand on. The answer is that none has been shown to, and the reason is worse for the vault’s position than a simple absence of studies.
  • Feed algorithms and attitudes in an election campaign is the largest randomized test of feed construction that exists: Facebook and Instagram users switched to a reverse-chronological feed for three months during the 2020 US election. The intervention worked as an intervention. Time on platform and activity fell substantially, and the content mix shifted several ways at once — more political and untrustworthy material, less incivility, more moderate sources on Facebook. Issue polarization, affective polarization, political knowledge, and other key attitudes showed no significant change. A change large enough to reshape exposure and behaviour moved the outcomes not at all.
  • Hilbert and colleagues’ pooled audit of 151 algorithmic audits points the same way about uniformity: roughly 8 to 10 percent of recommendations were “bad” while about a quarter actively protected users from self-induced harm. Recommenders in that accounting do more protecting than harming, which a purely engagement-maximizing model does not predict.
  • Three limits keep this from settling the question. The experiment studied adults, measured political rather than wellbeing outcomes, and ran three months, and the adolescent design-experiment base is still essentially empty. What it establishes is that the premise the two notes lean on is running ahead of its evidence, with one large well-powered result against it and none of comparable size in favour.
  • This is not an argument for age gates, and Less harmful social media now says so explicitly. If design changes do not move outcomes, that is evidence about the harm model itself rather than a reason to prefer exclusion; a gate would still have to show the effect redesign failed to show. The proportionality test survives. The confidence that redesign is waiting on the other side does not.
  • Corrected a claim added earlier the same day. Which social-media designs reduce harm for young users had cited Hilbert and colleagues’ pooled algorithm audits as showing recommenders protective more often than harmful, and concluded that this “is not the picture a purely engagement-maximizing model predicts.” That inference does not hold, and reading the paper’s method rather than its abstract shows why.
  • The audits are black-box audits of deployed systems. The authors explicitly decline to separate the ranking objective from the safety machinery layered on it, arguing by analogy to food safety that harm can be measured without opening the box. The pooled result therefore describes what platforms emit, not what engagement optimization does, and cannot distinguish emergent behaviour from deliberate safety intervention.
  • What the study coded as “good” is not crisis resourcing either. Its high-scoring audits are accurate influenza-vaccine pages in Google results, vaccine content on YouTube, and age-appropriate material on children’s channels. The protective pattern is a transition statistic: after a bad seed, most later recommendations are not bad, at 40 percent good-to-good against 11 percent bad-to-bad. The economical reading is regression toward popular generic content, which is what an engagement-maximizing recommender does by default, since mainstream material out-engages any niche. So the audits are consistent with engagement maximization rather than evidence against it, and they neither support nor undermine the design argument. Less harmful social media carries the same correction.
  • Researched Does better pornography production improve viewer outcomes, the supply-side branch of the harm-reduction alternative. The answer separates cleanly into a measured premise and an unmeasured conclusion.
  • Comparative content analysis is a mature literature. Fritz and Paul coded agentic and objectifying scripts across feminist, “for women,” and mainstream pornography; Sun and colleagues compared male- and female-directed work; Klaassen and Peter coded gender inequality in popular internet videos; and Fritz and colleagues compared depicted behaviours against a United States probability survey. The genre distinctions the reform argument depends on are measurable and have been measured. Sexual consent communication in best-selling pornography films shows the resolution available: consent communication is frequently depicted but mostly nonverbal, and the films model five identifiable scripts, including that explicit verbal consent is not natural.
  • That same study names the gap in its own conclusion, writing that further research is needed to establish whether viewers acquire, activate, or apply the scripts it documents. The literature best placed to close the gap leaves it open, and nothing found connects supply differences to viewer outcomes.
  • The cross-cutting result is the one worth recording. All three branches of the alternative in Alternatives to age verification — literacy, design, and production reform — have now been held to an outcome standard, and each returns a well-specified mechanism with no demonstrated effect on the outcome it targets. That is the objection the vault presses against age verification. It does not reverse the conclusion, since a gate has less evidence again plus privacy and autonomy costs the alternatives avoid, but the note now says plainly that these are the interventions whose mechanisms are best specified rather than the ones shown to work. Preferring a well-specified mechanism is defensible; presenting it as an evidence advantage is not, and the note had come close to doing so.
  • Corrected the framing of Does better pornography production improve viewer outcomes after the user pointed out that the viewer question is about content rather than production. The note had been asking two things at once, and the earlier edit made that worse by treating them as one branch. What is depicted — consent communication, agency, objectification, aggression, whose pleasure the camera follows — is what a viewer encounters and the only channel through which better material could change anyone. How it was made is invisible on screen and cannot be the mechanism, though it remains a real question on labour-rights grounds. Macleod’s finding is the hinge: viewers infer production ethics from an aesthetic, which staging can supply without the conditions behind it.
  • Reframing surfaced a partial answer the earlier pass had missed. The violent-versus-nonviolent distinction is a content property, visible on screen, and it does carry an outcome: Does pornography use cause adolescent sexual aggression found no association for nonviolent material and a weak one for violent material. So content type is not inert. What has no outcome attached is everything finer — agentic versus objectifying scripts, communicated versus assumed consent, representational range, mutual versus one-sided pleasure — which are exactly the distinctions the reform argument proposes. The one cut with evidence behind it is the crudest one available.
  • Created Pornography substitution hypothesis after the user asked whether violent pornography’s association with aggression runs from content to behaviour, from behaviour to content choice, or whether the material acts as an outlet that displaces offending. The vault held the first two as rivals and omitted the third, which has its own literature and points the other way.
  • The substitution evidence comes from availability shocks rather than from surveys. Kutchinsky found registered sex offences falling after Denmark’s 1965 liberalization, and later found that across the United States, Denmark, Sweden, and West Germany between 1964 and 1984 rape rose no faster than non-sexual violent crime in any of them. Diamond, Jozifkova, and Weiss found the same after 1989 in Czechia. Kendall exploited the arrival of the internet across United States states and found a ten-point rise in access tracking a roughly 7.3 percent fall in reported rape, with no comparable effect on other crime, concentrated in men aged 15 to 19 — which is the age band the age-verification debate is about. A German broadband study found substitution driven by child sexual abuse.
  • It is not settled, and two objections are decisive about that. Bhuller and colleagues found the opposite in Norway using the same kind of technology shock, so two well-designed European studies point opposite ways. And the ecological fallacy defeats aggregate inference in both directions, a point made most sharply by Ferguson and Hartley, who are otherwise sceptical of pornography harms; aggregate substitution licenses an individual-level claim no more than aggregate correlation licenses individual causation. Reported crime is also a contaminated outcome over these periods.
  • The Confluence model of sexual aggression reconciles the levels without requiring either literature to be wrong. If pornography raises risk only where other risk factors converge and displaces offending for the large majority where they do not, positive individual-level associations and negative population-level ones are what one should expect to see. On that account the main effect the vault has been asking after may not exist to be found, and the policy-relevant quantity is the interaction — harder to measure and much harder to legislate from, which may be why the debate keeps returning to a main effect.
  • Strengthened Pornography substitution hypothesis after reading Bhuller and colleagues’ Norwegian study directly rather than through a secondary account. The note had recorded Norway as a contrary result without saying what its authors concluded. They set out the same three mechanisms this note distinguishes — reporting, matching, and a direct effect on propensity — tested the reporting explanation against individual reporting data and found nothing supporting it, and concluded for the direct effect, “possibly as a result of increased consumption of pornography.” The published version finds increases in reports, charges, and convictions alike. Norway is therefore the strongest evidence against the substitution reading and cannot be set aside as a reporting artefact, because its authors looked for that and did not find it.
  • The note also now separates three claims that the word “substitution” runs together. Satiation is drive reduction and is temporary; habituation reduces response to a repeated stimulus, lasts longer, and is what the tolerance-escalation worry invokes; voluntary incapacitation needs no psychology at all, since time spent consuming is time not spent offending. Catharsis, the hydraulic model, is the weakest of the three and the one usually attacked, and neither of the others depends on it, so refuting catharsis leaves the substitution case untouched.
  • A Gemini-authored synthesis on the same subject was reviewed and not imported. Its distinctions were sound and are reflected above, but they are standard and did not need that source; its citation apparatus was unreliable, leaning on encyclopaedia entries, popular psychology sites, and revision notes, with several footnotes not supporting the claims attached to them; and its handling of Bhuller misattributed to those authors both a matching mechanism and a reporting explanation that the paper itself tests and rejects, which had the effect of neutralising the report’s strongest counter-evidence in favour of its own thesis. Nothing was committed, so nothing entered the evidence record.
  • Revised Pornography substitution hypothesis twice more under user challenge, and both challenges were right.
  • The first: the note had accepted Bhuller’s rejection of the reporting channel because the authors said they tested it, which is the same move — trusting a summary of a check rather than the check — that makes secondary accounts unreliable. Three grounds now qualify it. Gelman worked the reported coefficients against Norwegian base rates in 2011 and found the estimate implies the internet caused roughly half of all rapes and a third of child sexual abuse in the country, which is likelier to indicate an identification problem than a discovery. The Nordic paradox makes Norway the setting where a reporting confound deserves most scrutiny rather than least. And Diegmann’s German study, on the same treatment, found reporting effects real and adjusted for them. The note also no longer says the literature is lopsided: on rape specifically it is Kendall down, Diegmann null, Bhuller up, and the legalization studies are a different treatment that cannot be added to that tally.
  • The second challenge defeated the remaining defence, and the vault already held the argument. Convictions rising alongside reports looked like evidence against a pure reporting artefact. But a conviction is an institutional output rather than a measurement, which is Correction channel’s general claim, and for Nordic sexual offences there is specific evidence: Evidence patterns in rape and unlawful-threat cases found the same evidentiary pattern followed by conviction in 90 percent of the rape cases examined and 18 percent of the unlawful-threat cases, which the authors thought might reflect a lower operative threshold in rape. If that threshold moves with social attitude, reports, charges, and convictions are not three independent measurements but three outputs of one system whose sensitivity was changing across exactly Bhuller’s years.
  • This connects two clusters that had never been linked. The rättssäkerhet material and the pornography-evidence material share almost no vocabulary, so neither just links nor the structure audit surfaced the relation, and it took a question from the user to find it. The general lesson is Correction channel’s: when an outcome variable is produced by the institution whose behaviour is under study, it cannot arbitrate a dispute about that behaviour.

2026-07-28

  • Audited the vault’s use of rättssäkerhet after the user asked whether it was correct. The word appears in roughly fifty notes, titles the criminal-justice cluster’s central synthesis, and was defined nowhere: Peczenik, Frändberg, Vahlne Westerhäll, and the constitutional mandate were all absent, while Two Visions of Time and Per Olof Ekelöf had given the sibling term rättsstat a home. A load-bearing premise the vault restated in fifty places and never linked is exactly the shape just distill exists to find, and no command reported it.
  • Wrote Rättssäkerhet as the concept note, covering the agreed core of predictability plus power under law, the formal and material axis from Ekelöf 1942 through Axberger 1988 to Peczenik 1986 and Vahlne Westerhäll 2002, Frändberg’s objection that the split is drawn pejoratively, the traditional and “modern” axis over whether protection from crime is itself a rättssäkerhet requirement, and the three translation registers. Status working: Frändberg’s and Peczenik’s primary texts are characterized from summaries and Zila’s citations, which is a named gap rather than youth.
  • Imported Josef Zila, “Om rättssäkerhet,” Svensk Juristtidning 1990 s. 284–305, and wrote its source note Om rättssäkerhet. It is the grounding artifact for the concept note: a survey that treats the term’s vagueness as its object, records that 8 kap. 22 § RF names rättssäkerhetens krav while saying nothing about its meaning, and routes to the Swedish primary literature.
  • Corrected two mistranslations to “legal certainty.” Rättssäkerheten i brottmål had Justitiekanslern “defining legal certainty” as confidence that nobody is convicted without proof beyond reasonable doubt, which is a category error in English: legal certainty means foreseeability, not outcome accuracy. Dropped the same alias from Rättssäkerhet in Swedish criminal cases and gave that note an opening paragraph naming which of the three senses it argues in, since its epistemic claim goes a step past even the JK sense.
  • Renamed a section in Reliability of mobile forensic extraction as evidence from “What would make an extraction rättssäker” to “What would make an extraction examinable.” The section answers with reliability engineering — pre-exploitation imaging, verification without trusting the first tool — and no register of the term makes a method rättssäker.
  • Narrowed the disarmament claim in Scandinavian legal realism and Swedish lay judges and political influence. Both said the school removed the vocabulary in which state power could be called wrong as a matter of principle, and that the rättssäkerhet register arrived only with the post-ECHR paradigm shift. That is false as stated: Ekelöf published “Rättssamhälle och rättssäkerhet” in Svensk Juristtidning in 1942, the constitution has directed Lagrådet at rättssäkerhetens krav since 1979, and Peczenik was arguing the material case by 1986 — all before the 1995 Convention incorporation. The surviving claim is stronger than the one replaced: realism did not remove the word but fixed it in a form holding its value “oberoende av den gällande rättens innehåll,” and a term defined that way cannot say a law’s content is wrong. The material turn that re-arms it is Swedish legal theory’s own work, not a Strasbourg import. Added the 1942 article to Per Olof Ekelöf, where his own writing on the concept had been missing.
  • Added the traditional-versus-modern axis to Anonyma vittnen. Both sides of that reform claim the word, and the split — protection against the state against protection against crime, which the riksdag’s committee on justice rejected in 1985 as rättstrygghet rather than rättssäkerhet — is the axis the dispute actually runs on.
  • Repaired two broken source artifacts. Dömda’s Storytel EPUB carried 38 trailing NUL bytes after </package> in content.opf, which strict XML parsers reject; the padding was stripped and the archive repacked with no content files altered. The artifact recorded as 2007-gilbert-the-disposable-male.epub turned out to be a PDF with the wrong extension; it was renamed to .pdf without changing any bytes and the recorded SHA-256 still verifies. The genuine EPUB of the same book was then added alongside, so The Disposable Male now bundles both formats with their hashes.
  • Greater male variability and Male variability, education, and life outcomes updated their sources references to the renamed PDF.
  • Measured F-Droid’s developer-signed coverage directly instead of repeating the project’s own headline figure. Intersecting the published index-v2 main-repo index with AllowedAPKSigningKeys in fdroiddata metadata, and checking each app’s latest published APK against its pinned developer certificate, gives 1,339 of 4,113 apps — 32.6% — on 2026-07-28, against the roughly 21% F-Droid reported for end-2025. Android app distribution trust models, F-Droid, Reproducible builds, and PrivSec.dev on F-Droid security issues all carried the stale fifth and now carry the counted share with its method in a footnote.
  • Created Play App Signing, a premise the F-Droid custody argument had been leaning on without ever naming. Google generates or ingests the app signing key, holds it in Cloud KMS, will not return a copy, and has required this of every app created since August 2021, so distributor-held keys are the majority arrangement on Android. PrivSec.dev on F-Droid security issues concedes the parity in its own text, which undercuts the trusted-party count it argues from; Android app distribution trust models now records that concession.
  • Sharpened the custody analysis in Android app distribution trust models. A compromised F-Droid keystore reaches only F-Droid-signed apps, not the developer-signed third of the repository, because F-Droid holds no key for those. Separated the two senses of “reproducible” that were blurred: the developer-signed publishing path removes the key from F-Droid, while the verification server’s per-app reproducibility status audits F-Droid’s build honesty and leaves the APK F-Droid-signed. Added the counterweight that developer-held keys fail often — 26 of 1,105 IzzyOnDroid apps had changed signing key by 2023, none maliciously — and the limitation that a rebuild guarantee discharges only when someone runs the rebuild.
  • Imported Understanding Build Reproducibility in the F-Droid Ecosystem, an ACM REP ‘26 study by Nanni, Malka, Zacchiroli, Zimmermann, and D’Angelo, preserved as sources/2026-07-28-fdroid-build-reproducibility-acm-rep-2026.pdf. It is the first empirical measurement of whether a distribution’s historically reproducible artifacts can still be rebuilt: of 18,904 F-Droid versions that passed a check when published, 83.7% still rebuild, missing dependencies cause about 76% of the failures, and 94% of the successful rebuilds are still bitwise identical.
  • Extracted Rebuildability from that study. The vault had been treating reproducibility as one property, and the paper’s central result only parses once the capacity to run a build is separated from the determinism of its output; the two fail for unrelated reasons and only the second is an engineering problem the project controls.
  • Placed the finding rather than filing it. Source-to-binary correspondence now says the guarantee is a dated audit rather than a standing property, which is a material limitation the note lacked. Reproducible builds gained the result that determinism is not the binding constraint. Android app distribution trust models gained the two ways its reproducibility backstop is weaker than it looks: F-Droid’s verification server is run by F-Droid, so it does not address a compromise of F-Droid’s own infrastructure, and independent instances remain an undeployed goal; and the check itself expires as rebuildability decays. The open question about rebuilder uptake was rewritten to track that deployment instead.
  • Extended the desktop security cluster to five systems and fixed what the previous pass left behind. Tails was missing entirely while Qubes OS, Whonix, Kicksecure, and secureblue all had notes, which had distorted the comparison: Kicksecure claimed its live mode was unique “among the three compared,” a statement true only because the canonical amnesic system was not among them.
  • Corrected two errors in the Qubes advisory-record citation. The claim was dated to 2025 when the paper was posted in July 2026 and analyzes a 2011–2025 record, and the transient-execution figure of 21% was attached to post-2018 advisories when 21.1% is the share of all 109 bulletins and the post-2018 share is 31.5%. The paper was cited in two notes by bare arXiv URL with no source note and no preserved artifact; it is now Qubes OS Security in the Public Record with a local PDF.
  • Named the epistemic limit that citation kept crossing. The paper measures the advisory record, not the defect population, and says so; Correction channel already carried the general premise that a channel’s own outputs cannot establish its own error rate. The two regions of the vault were unconnected, so the same reasoning was being rediscovered in a security context after having been worked out in a criminal-justice one.
  • Added reproducibility as a comparison axis. Tails has shipped reproducible images since 3.3 in 2017, and no other system in the set documents a procedure by which a user could perform the equivalent check. That is the only row in the comparison an outsider can verify; every other claim rests on believing a project’s own documentation. The note records the absence of a published check rather than the absence of the property, which is the stronger claim the evidence does not support.
  • Treated competitor comparisons as arguments, not findings. Whonix comparison with other anonymity systems is preserved because its architectural claim about Tor client placement is sound, but its feature rows evaluate Tails 2.4 from 2016 and record Tails as lacking deterministic builds, which stopped being true a year after that version shipped. The failure mode is decay rather than invention: a project maintains its own rows and lets a rival’s rot.
  • Extracted Evil maid attack, a premise three notes in the cluster restated without linking. Qubes ships Anti Evil Maid, Kicksecure advertises deep-scan readiness, and Tails documents the firmware exposure it cannot address; none of them named the shared attack. Writing it surfaced a constraint none of the three notes carried: Qubes AEM requires Intel TXT and is therefore unavailable on a machine whose Management Engine has been neutralized for coreboot, so open firmware and this defense are alternatives rather than layers.
  • Repaired abbreviation coverage across all nine cluster notes. Dozens of terms were used without definition — IOMMU, VT-x, SLSA, KSPP, SUID, DNSSEC, TCP, and others — while several notes defined abbreviations they never used. Lint does not catch either case, so the gap had been invisible to the checks.
  • Added CalyxOS and its hiatus record. The vault had no note on it at all despite covering GrapheneOS in depth, which left the mobile comparison with no account of the other de-Googled Android. It is treated in its own section rather than as a fifth column, because it is a software choice on top of the same hardware and every hardware row would duplicate the stock device’s.
  • Recorded the eleven-month update gap as the central fact. CalyxOS suspended releases in August 2025 after its founder and technical lead departed, and resumed on 1 July 2026. The primary artifact carries the detail that matters: the project recommended that users uninstall the operating system, pushed a final over-the-air update to reach people who had missed the announcement, and estimated four to six months for work that took about eleven. CalyxOS hiatus and return announcements preserves both pages.
  • Separated the technical comparison from the advocacy around it. microG’s privileged domain and signature spoofing are verifiable from microG’s own documentation, whose developers concede the security effect while disputing the stronger claim that it breaks Android’s signature model. Much of the circulating “CalyxOS is insecure” material originates in the GrapheneOS community, so the note traces claims to mechanisms rather than to advocates — the same discipline applied earlier today to Whonix comparison with other anonymity systems.
  • Date-scoped rather than inherited the standard criticisms. The detailed Privacy Guides comparison is from 2022 and its publisher flags it as over four years old, so the eSIM, F-Droid privileged extension, and user-profile points are recorded as unverified for the 7.x series rather than restated as current fact.
  • Opened Does CalyxOS sustain reliable security updates after its hiatus. Nothing available in July 2026 can answer it, since only elapsed time will, and the decision it bears on is concrete: whether CalyxOS is defensible for someone who cannot obtain Pixel hardware and whose real alternative is stock vendor Android rather than GrapheneOS. Twelve monthly releases checked against Android bulletin dates would settle it.
  • Repaired pre-existing abbreviation debt in Mobile operating system security comparison while editing it: CPU, DMA, EU, FBI, LTE, LTS, MiB, ROM, and USB were used undefined, and PIN was defined but unused.
  • Ran the scholarly sweep the desktop cluster never got, and recorded the negative result. OpenAlex across five query framings, plus a federated search of arXiv, Semantic Scholar, DBLP, and CORE, returned no independent security evaluation of Qubes, Tails, Whonix, Kicksecure, or secureblue. A title search for Tails returned nothing; a title search for “Qubes” returned 227 works, none about this operating system. Desktop operating system security comparison now states this, which converts its reliance on project self-documentation from an apparent shortcut into the actual state of the evidence. The sweep excluded Google Scholar and a manual conference-proceedings pass. That scoping lives here rather than in the note, along with the engines and query framings: a reader of a desktop-OS comparison needs the finding, not the method that produced it.
  • Fixed the instruction defect that caused the earlier tooling failure. AGENTS.md was precise about local search — just search, just semantic, rg, ugrep, rga — and vague about external search, offering only “the best available outside context,” which is why a session that read it still defaulted to the harness web-search tool and collected search-optimized listicles. It now names the Exa server and the scholarly servers as the default route, and points at the tested-tooling reference.
  • Rewrote the deep-research reference triggers as workflow positions. They had been written as felt states — load when “unfamiliar” with a search surface, or when the first wave “repeats itself” — which fail because an agent does not notice being unfamiliar; it simply uses the first tool at hand. Two references are now read on schedule (tooling before the first external search, evidence and synthesis before writing), and the skill opens by creating tasks for the stages most often skipped without leaving a trace, the tool inventory and the disconfirmation wave.
  • Benchmarked the semantic search that shipped this morning and changed its default. Twelve queries with known targets across three engines: keyword found the right note in the top five 4 times, pure vector 7, and the hybrid default 5. Fusion rescued one vector miss while losing three vector hits outright and demoting two more, and the losses landed on paraphrase and cross-lingual queries — the cases semantic search exists to answer. just semantic is now purely vector-ranked, with --hybrid available for a query mixing a concept with a rare literal token.
  • Corrected two claims the benchmark contradicted. Vector ranking handles citation identifiers and foreign terms well: C-136/17, SOU 2024:75, and utgivningsbevis each returned the right note first, so the warning about embeddings missing case numbers was wrong. What it does miss is the rare literal token — sdwdate, onion-grater, a personal name used once — which keyword search finds immediately.
  • Recorded that a stale index fails silently. The first benchmark run scored three queries as retrieval failures that were only notes created earlier in the session and never embedded. A missing note is indistinguishable from an absent one, which matters because the vault uses these searches to conclude that something is not there.
  • Retired the claim that no command finds restated premises. AGENTS.md and the distill-syntheses skill both asserted that a premise restated in different words is invisible to search, citing the four notes that argued an institution cannot measure its own error rate while sharing almost no vocabulary. Querying that premise as a sentence now returns all of them, plus two the passage did not know about. Both files now name just semantic as the first step of an extraction pass, and both keep the point that it aims the reading rather than replacing it: the query has to be guessed before it can be run, so it finds premises already suspected and stays silent about the rest.
  • Rebalanced the discovery guidance. AGENTS.md had presented just search as the discovery tool and semantic search as a complement sixteen lines later, and no skill mentioned just semantic at all. Discovery is now stated as two searches whose results barely overlap.
  • Routed semantic search into the seven remaining skills that do discovery. Each got the use its own pass actually needs, rather than a mention of the command: audit-structure queries a cluster’s central claim to find duplication the link graph cannot surface, since the defect is precisely that nothing links the pair; triage-wanted-pages checks whether a redlink’s subject already exists under a handle its author never considered, which no keyword search on the redlink text can find; import-source runs placement as a meaning query, because the notes a source bears on are usually the ones written in vocabulary the source never uses; research-questions queries the question as a claim, since backlinks show what cites a question rather than what bears on it; maintain-tags finds synonym pairs and under-applied tags, both of which are meaning failures while every section of just tags is spelling or co-occurrence; maintain-index-note checks for a duplicate hub, which by construction does not share the new hub’s vocabulary; and review-note-lifecycle re-queries a corrected claim, because review_after is set per note while a stale fact is not, so a superseded price or rule survives in notes whose review is not due.
  • Left maintain-graphs-note alone. It edits Mermaid markers and generated graphs and has no discovery step, and adding the command there would be decoration. Each addition states its limits: the query must be guessed before it can be run, so semantic search confirms what a pass already suspects and stays silent about the rest.
  • Imported two YouTube videos on Pokemon Go and Niantic, saving timestamped transcripts as sources/2026-07-28-patrick-cc-why-the-cia-created-pokemon-go-timestamped-transcript.txt and sources/2026-07-28-infographics-show-you-didnt-play-pokemon-go-timestamped-transcript.txt, with source notes Why The CIA Created Pokémon Go and You Didn’t Play Pokemon GO. Both argue that the game was a covert mapping operation. The research pass found the underlying corporate history accurate and the espionage thesis unsupported, so both notes are written as assessments rather than summaries, with per-claim timestamps so a later citation can point at a checked segment instead of the video.
  • The placement pass found the vault held nothing on any of it: no Niantic, no augmented reality, no visual positioning, no geospatial AI. Created Niantic Spatial as the entity, covering the Keyhole-to-Google-to-Niantic lineage, the In-Q-Tel investment of February 2003 on behalf of NIMA, the 2025 Scopely split, and the December 2025 Vantor defense partnership. Primary sources carried the load here: In-Q-Tel’s own 2003 press release, Niantic’s 2024 large geospatial model post, Scopely’s acquisition release, and the joint Niantic Spatial and Vantor announcement.
  • Extracted two premises both videos argue from without stating. Visual positioning systems defines the mechanism as a category, written at its own scope rather than Niantic’s, and is where the videos’ tracking and targeting claims actually fail: a VPS localizes a consenting device against a reference map and is not a channel for finding a person. Training data provenance carries the reason the central dispute cannot be settled — once scans are absorbed into model weights, neither the accusation nor the denial can be checked from outside.
  • Opened What happens to data consent when a company is split or sold as a question note. The vault now holds two instances of the same shape, Niantic and 23andMe, and no answer; both notes now link it, and Data minimization gained a paragraph on the future acquirer as the case minimization handles worst.
  • Recorded the disconfirming evidence rather than the thesis: Foreign Policy’s 2024 reporting, drawing on seven former NSA, CIA, and Energy Department officials, found no evidence Pokemon Go was ever used for intelligence collection, and describes US agencies treating it as a counterintelligence nuisance. It also supplies the one real institutional link neither video mentions, that In-Q-Tel’s founding chief executive sat on Niantic’s board.
  • Cited Trouw, Ars Technica, the Guardian, Euronews, and IGN directly rather than extracting concepts from them. They report one unfolding dispute and support claims in Niantic Spatial about who said what and when, which is source work rather than a reusable premise. Left In-Q-Tel and Vantor as redlinks: both are independently nameable and both recur across these notes, but neither was researched beyond its role here.
  • Cut the research method out of Desktop operating system security comparison. The note had carried the engines queried, the number of query framings, and the comic near-misses a title search for “Qubes” returns. That is a record of how a session worked, not a fact about desktop operating systems, and AGENTS.md already routes research-state commentary here. What survives is the finding a reader needs: no independent evaluation exists, so every comparative claim rests on documentation written by the party it flatters.
  • Queued the per-note provenance question in Todo. Cutting the sweep method out of the desktop comparison was right for the reader and lost what makes a negative finding checkable, since Log.md is chronological rather than indexed by note. The item records the three things to settle before adding a field — whether Log.md backlinks already reach the note, whether the field would render, and what it is actually for — and holds it against the bar AGENTS.md set after the subtype episode, which one instance does not clear.
  • Answered whether Qubes is the most secure desktop, which it is not unconditionally, and gave Desktop operating system security comparison the frame it lacked. Isolation and hardening reduce different terms in the same risk: hardening lowers the chance an exploit lands, compartmentalization lowers what it reaches. Compartmentalization helps least when the asset sits inside the compromised unit, which is the browser case — sessions, cookies, and saved passwords are taken whether or not the attacker leaves the qube. Qubes’ real answers are configuration rather than architecture: splitting by activity, and disposable qubes, which bound a compromise in time as hardening cannot.
  • Imported the evidence that Xen isolation has failed in practice. Xen isolation failures affecting Qubes OS pairs Project Zero’s XSA-212 write-up, whose exploit ran commands as root in every other paravirtualized guest including dom0, with QSB-029, in which Qubes states that such a bug lets an attacker who has already compromised a browser “compromise a whole Qubes system.” That is the chain compartmentalization exists to break, described by the project itself. Qubes dropped paravirtualization in 4.0 in response, so the documents evidence a failure that happened rather than today’s attack surface; what the fix did was move the guarantee from Xen’s software into the processor, where 23 of 73 post-2018 bulletins live and where software cannot fully fix it.
  • Ranked secureblue against Kicksecure on exploit mitigation, where secureblue wins: hardened_malloc globally, SELinux rather than AppArmor, and a hardened Chromium. Kicksecure’s deprecation of the allocator in November 2025 is a reasoned trade, not neglect — it cites stability, breakage of Firefox, Tor Browser, and Chromium, and AMD64-only support against its ARM64 and PPC goals — and secureblue can carry it partly because it builds its own browser. The ordering inverts on anti-forensics and physical custody, where Kicksecure has RAM wipe, default encryption, live mode, a panic key, and a warrant canary. The note records that the vault holds only Kicksecure’s account of both projects.
  • Corrected the claim that a Qubes user can simply install a hardened browser. The line was true but unchecked, and checking changed what it means. Trivalent is not locked to secureblue: it is a Fedora package from repo.secureblue.dev, built for Fedora 42 to 44 on x86_64 and aarch64, so it installs in a stock Qubes Fedora template. What does not come with it is the rest of secureblue — global hardened_malloc, the sudoless model, the SUID reduction — and the project documents the browser as intended for use with the allocator it does not get there.
  • Ran down the secureblue-template question and found no roadmap to track. QubesOS/qubes-issues#8997, “Wayland proxy compositor,” has been open since February 2024 with default priority, no milestone, and no update since January 2025; QubesOS/qubes-gui-agent-wayland has one contributor and no push since October 2023. The issue also shows the motivation is GPU support rather than compatibility with other distributions, so secureblue-in-Qubes is not a scheduled outcome. Qubes forum discussion adds the non-technical half: someone wrote a Kicksecure template and nobody wrote a secureblue one.
  • Recorded that the two hardening stacks cannot be combined in one qube. Trivalent is packaged for the Fedora family only and Kicksecure is Debian, so the choice is a hardened browser on a baseline template against a hardened template with a weaker browser. Inside Qubes the browser is the better place to spend, because Qubes already supplies the containment that template hardening partly duplicates and supplies nothing that lowers the odds of the browser falling. The Qubes-shaped answer is per-qube: Fedora with Trivalent for browsing, Kicksecure templates where the risk is escalation rather than hostile web content.
  • Separated the privacy job of compartmentalization from the security one in Desktop operating system security comparison. Bounding a compromise and keeping identities unlinked are different problems, and Qubes is stronger at the second for reasons unrelated to exploits. A persona is not only its cookies: the source address, the browser fingerprint, and the timing of when two accounts are active all sit below the browser, so only a boundary that carries its own network path separates them. That ranks browser profiles below operating-system profiles below a qube with its own network VM, and it is the strongest form of the claim that Qubes is a privacy tool as much as a security one.
  • Answered what Trivalent provides for identity separation. Chromium profiles are available. Firefox-style per-tab containers are not, and the gap is structural: Chromium has no native contextual identity, so on Chromium the feature is an extension, and Trivalent disables extensions by default and permits only Manifest V3. What it does instead is partition — third-party cookies blocked, split cache, partitioned connections, strict origin and origin-keyed process isolation, sandboxed iframes isolated per site — plus its own preferences for launching into Incognito. The note records why that is not the same thing: partitioning stops one site correlating you with another, and does nothing about one site linking two accounts you hold there, because that runs through address, fingerprint, and timing.

2026-07-29

  • Split the two-part title on Whonix comparison with other anonymity systems into title and subtitle. The Whonix wiki page prints “Anonymity Operating System Comparison - Whonix vs Tails vs Tor Browser Bundle”, and the note carried the whole string in title, so the separator rendered as a hyphen among the hyphens that join words rather than as the dash the page means. Both halves keep the title case the work prints. Swept the other source notes for the same defect and found none.
  • Structural audit across 27 clusters and 513 subject notes. Checked orphans, dead ends, clusters, bridges, type warranties, and cross-cluster paraphrase via semantic search. One finding: Duress credentials and coercive extraction cited four source notes about Swedish remand isolation but never wikilinked the synthesis note Swedish remand detention and restrictions that integrates them; added the link. Orphans are infrastructure source cards linked only from meta notes (expected). Dead ends are all healthy source notes absorbing links. No type-warranty, duplication, or concentration issues found. The vault graph is in good shape.
  • Distillation pass extracted Journalistic purposes under Article 85 GDPR as a new concept note. Six notes were restating the same four-part CJEU test in different words and linking to the source case Legal Newsdesk Sweden C-199-24 rather than to a shared concept page. Rewrote the journalism-test passages in Naming convicted people in Swedish media, Dumpen, Utgivningsbevis and people-search services, EU and Swedish data removal rights, Swedish public-record privacy, and Unmasking pseudonymous public figures to link the new concept at the point where the test is invoked. The concept note is extracted from the CJEU judgment rather than from the citing sentences that compressed it, preserving the processing-specific structure, the four indicators as criteria rather than a checklist, and the consequences of failing the test.
  • Connected two notes to the existing Correction channel concept: What is the production false-positive rate of automated CSAM detection and Operationalizing content moderation accuracy in the DSA both argued that an institution’s own outputs cannot calibrate its accuracy without linking the note that formalizes that principle.
  • Read the five “leaning on direct citation” notes flagged by just distill. Three are correctly typed: Greater male variability and Invoking the Fifth Amendment can itself be used as evidence are self-contained concepts whose source cards supply the factual base, and Pornography literacy and harm reduction is a synthesis whose evidence necessarily sits in source cards. Age-verification circumvention services is well-linked to Displacement, VPN service legal risk, and Alternatives to age verification. Naming convicted people in Swedish media was the only one whose restated premises called for extraction, and the journalism-test concept above addresses it.
  • Tag maintenance pass across 285 tags and 518 notes. Retired academic-paper (format descriptor, not navigable). Expanded 7 under-applied singletons: DSA to Operationalizing content moderation accuracy in the DSA; search-indexing to Challenging name-search indexing of criminal coverage, Dumpen, and Naming convicted people in Swedish media; measurement to Is the gender-equality paradox a measurement artifact; prisons to Double occupancy in Swedish remand prisons; child-protection to Virtual private networks and child protection EPRS; retail to Tacticalstore (tactical-retail is a subtype). Fixed two near-miss containment gaps: added privacy to Journalistic purposes under Article 85 GDPR (GDPR note missing its parent privacy tag) and Desktop operating system security comparison (had network-privacy without privacy). Checked all 89 singletons and every near-miss row; preserved competitor/competition distinction (entity vs topic), detention/remand-detention (general vs Swedish pre-trial), and sexual-learning/sexual-education (informal vs formal). No merges, aliases, or renames needed. The vocabulary is clean.
  • Index maintenance pass: added three missing top-level entries for synthesis notes that carry their own substantial reading paths but were not reachable from any existing Index entry — Desktop operating system security comparison, Secure and decentralized communication stacks, and Cryptocurrency design comparison. All three are self-contained evaluation domains with extensive onward navigation into their entity and concept notes, and none is linked from the Privacy venture atlas or the other existing entries. Reordered the topic list so the three technology-evaluation entries sit near the Privacy venture atlas, keeping the rough thematic flow from venture tools through Swedish law to money.
  • Second Index pass added Mobile operating system security comparison beside the desktop entry. Its 25-note cluster was reachable only through the “See also” list at the bottom of Desktop operating system security comparison, which is not where a reader choosing a phone platform would look. The two notes are peers rather than a hub and a subordinate, and each answers a distinct starting intention. Checked the other unlisted clusters and left them alone: the private-AI cluster routes through Privacy venture atlas, Scandinavian legal realism has a prose reading path from Rättssäkerhet in Swedish criminal cases, and the remaining clusters are too narrow for a top-level entry.
  • Third Index pass, run against the current maintain-index-note skill after finding the .claude/skills/ mirror two commits stale and resyncing it with just sync-skills. The canonical skill requires an entry table, a cluster sweep, a grouping threshold, and a report on every type: index page; the stale mirror carried none of them, which is why the two earlier passes today stopped at adding entries.
  • Grouped the topic list under four headings — “Choosing tools”, “Building a privacy business”, “Swedish law and the state”, and “Society and policy” — since thirteen ungrouped entries had passed the point where a reader scans the list instead of choosing from it.
  • Added Scandinavian legal realism as a topic entry. The earlier pass left it out because a prose path reaches it from Rättssäkerhet in Swedish criminal cases, but reachability is not the test: a reader arriving for the Uppsala school and value nihilism does not self-sort into a criminal-procedure entry, and the cluster is 22 notes behind a hub with 23 outbound links.
  • Created Minors and online regulation as a type: index page and routed EU age verification and Internet privacy and Pornography literacy and harm reduction behind it. The cluster is 69 notes, the vault’s second largest, and its two Index entries covered two of its four sub-domains: the automated-detection material around Automated CSAM detection and CSAM scanning externalizes error costs, and the platform-design material around Less harmful social media, had no top-level door and sat two hops behind glosses that did not mention them. The new page routes by which intervention a note examines — identifying the user, constraining design, teaching the reader, detecting abuse material — and leaves the comparison between them to Alternatives to age verification, which argues that case and is not an index.
  • Connected Male variability, education, and life outcomes to the six source notes it discussed in bare prose: The Myth of Male Power, The Disposable Male, What if boys were never the problem, Some cold facts about math and gender, Overdiagnosis of ADHD in children and adolescents, and Deconstructing the gender-equality paradox. All six were already in its sources frontmatter, so the provenance was recorded and only the links were missing. Five outbound links over a ten-note cluster was the thinnest reading path of any listed entry, and the Index promises a path the destination could not deliver.
  • Added Privacy gear market scan to Privacy venture atlas. It is the second-most-linked note in the vault’s largest cluster, and the atlas routed to the Privacy gear webshop proposal without naming the research behind it.
  • Left Swedish business tax map alone and recorded why. Its F-tax, limited-company, and sole-trader sections carry definitional prose where an index page would normally carry links, which usually means the destination notes were never written. Here no other note in the vault mentions F-skatt, aktiebolag taxation, or enskild näringsverksamhet, so there is one consumer rather than a suppressed queue, and writing the three notes needs Skatteverket rates verified under a review schedule rather than a navigation pass.
  • Cluster sweep outcomes for the remaining neighborhoods above ten notes: cluster 1 (privacy competitors, 90), cluster 4 (private AI, 38), and cluster 5 (case for privacy, 37) route through Privacy venture atlas at one hop, verified against their hubs. Cluster 6 (people-search, 28) routes through Swedish public-record privacy, whose center Utgivningsbevis and people-search services sits one hop behind it. Clusters 3, 7, 9, 10, 11, 12, and 13 have their centers listed. Everything below ten notes was left to Graphs, Tags, search, and backlinks.
  • Found two orphaned source cards outside this pass’s scope: ChatGPT Sites official publishing documentation and Cloudflare Pages and Access official publishing documentation were imported for Publishing the wiki privately and are linked from nothing.
  • That orphan finding was a false positive. Both cards are linked from Publishing the wiki privately, which is a meta note, and just links excludes index and meta notes from the graph, so a source card whose only legitimate citer is a meta note always reports as an orphan. No subject note has business citing publishing-platform documentation, so the correct action was none.
  • Renamed the online child protection index to Minors and online regulation. The old handle took its name from the regulatory sector, and the page routes to Pornography literacy and harm reduction, Better pornography production, and the open questions on adolescent outcomes — notes arguing that the protection framing is what is under test. Naming the region for one answer to its own open question put that answer at the wiki’s most visible surface, against the analytical-voice rule on inherited framing and the rule against treating minors as one category. The regulatory vocabulary moved to aliases, where it still serves discovery.
  • Split Scandinavian legal realism in two. The note had grown a thesis it could not carry as a concept: it defined the school for twenty-one backlinks and also argued the wiki’s own conclusion about Swedish criminal procedure. The definition, the “might makes right” controversy, the published criticisms, and the documented channels stay in the concept note, now stated as attributed reception rather than the wiki’s verdict. The argument moved to Realist foundations of Swedish rättssäkerhet deficits, a synthesis whose thesis is that the doctrine underwrites the mechanisms without having caused them.
  • The synthesis is a three-channel spine rather than the earlier single claim, which was the user’s correction: rights-rejection is the primary channel but not the only one, since the lay-judges argument runs through validity-as-observance and Per Olof Ekelöf’s rejection of rights-protection as procedure’s function is a documented doctrinal position, not a homology.
  • A disconfirmation pass tested whether channel one is a civil-law commonplace. It is not. Weigend’s German chapter records absolute exclusion for intimate conversations captured in surveillance, a balancing test otherwise, and routine exclusion for intentional violations, in a system that emphasizes substantive truth at least as heavily as Sweden’s; Turner and Weigend record exclusionary rules expanding across both procedural families through the twentieth century. Sweden’s position is a choice within the civil-law family, not a feature of it. Free evaluation of evidence in Sweden now carries the comparison.
  • Imported two Swedish sources that changed the argument rather than decorating it. Något om Ekelöfs teleologiska metod (Klami, SvJT 1990) attributes Ekelöf’s method to the school outright and records that rättssäkerhet was an important teleological argument for him, which forced the disarmament claim to be restated: the term was operative and confined to its formal register, not missing. Hägerström och juridiken (Frändberg, SvJT 2011) calls the received account of Hägerström’s influence on jurists a set of usual clichés, and is carried as a limit on the synthesis rather than as support for it.
  • Strang’s 2009 “Two Generations” bounds every influence claim in both notes: Lundstedt and Olivecrona aimed at transforming law into a vehicle for reform, while Ross and Hedenius worked to blunt that radicalism, so a claim has to name which generation it runs through. The Helsinki repository blocked retrieval through three escalation rungs, so it is cited without a local artifact.
  • Repointed the notes that wanted the argument rather than the definition: Swedish lay judges and political influence, Free evaluation of evidence in Sweden, Rättssäkerhet, Per Olof Ekelöf, and Did legal realism weaken Swedish rights protection, which keeps the causal question the synthesis explicitly declines to close.
  • Ran an Index pass after the realism split. The entry gloss had gone stale: it promised “their reach into Nordic legal institutions” from a note that no longer carries that argument. Scandinavian legal realism stays the door because it is still the most-linked note in its own 26-note cluster, and it gained a reading path gathering the six people notes, the five source notes, and the two argument notes the prose reached only in passing. Realist foundations of Swedish rättssäkerhet deficits is the Index entry, replacing the concept note that held the row before. The user’s objection settled it: the vault’s product is its warranted conclusions, and the Index reflected that everywhere except this row — nine syntheses, three index pages, and one concept. “What was this school” is an encyclopedia lookup that search, Tags, and backlinks already serve, so it does not earn a door. Rättssäkerhet in Swedish criminal cases also routes to the synthesis, and the two entries under the Swedish heading hold distinct intentions: what is wrong with how Sweden convicts people, and why the procedure is built that way.
  • The cluster sweep recorded an outcome for every neighborhood above ten notes. Privacy competitors (90), age verification (69), private AI (38), and the privacy case (37) route through Privacy venture atlas and Minors and online regulation, both verified to link their cluster’s center. The criminal-procedure cluster (63) carries two doors, Rättssäkerhet in Swedish criminal cases and Desktop operating system security comparison, which survive the intention test as genuinely different arrivals. In the people-search cluster (27) the listed door is not the most-linked note, and it stays that way: Swedish public-record privacy is what a reader arrives wanting, while Utgivningsbevis and people-search services names a mechanism.
  • Recorded a finding about Swedish business tax map so the next pass does not re-derive it. It is the thinnest index page in the vault, four outbound links against the atlas’s forty-six, because four of its sections carry definitional prose where a routing page would carry links. Both just search and just semantic found no other note restating F-tax, limited-company, or sole-trader taxation, so those would land as single-backlink notes. They belong in the redlink queue rather than in an index pass.
  • Ran the Skatteverket pass anyway, on the separate argument that the tax map’s review_after was scheduling a revalidation of 2026 rates held in the prose of a note typed index. Wrote F-skatt, Preliminärskatt, Arbetsgivaravgifter, Egenavgifter, and Utdelning i fåmansföretag, each carrying its own rates, citations, and review date, and collapsed the three link-free sections of Swedish business tax map into routing. Every section of the map now links out; the only claim it still makes is the 20.6-percent corporate rate that makes the owner layer a separate question.
  • Preserved six pages under src/sources/ behind two cards. Skatteverket business tax pages 2026 bundles the five agency pages; Verksamt.se guidance on egenavgifter is separate because verksamt.se is a different issuer, and it is the only source carrying the sole trader’s general contribution reduction.
  • The closely held company rules changed materially from income year 2026, which no earlier pass had noticed. The simplification rule was abolished and replaced by a single base amount of four inkomstbasbelopp, 322,400 SEK for 2026. The budget proposition still describes the old 2.75 rule, so reading the legislation rather than Skatteverket’s operational page understates a sole owner’s dividend allowance by about 100,000 SEK. Both source cards record that divergence.
  • Two dates now drive the review schedule rather than a round year. The temporary employer contribution for 19–23-year-olds runs to 30 September 2027 and is legislated to end, and ordinary rates are set each autumn for the following January, so the rate-bearing notes carry review_after: 2026-12-15 instead of the map’s former 2027-01-15.
  • Corrected the new synthesis’s title and stripped editing voice from its prose. The title had been set to the claim as a sentence while the handle carried the name the user chose; the name is now the title, and the sentence moved to aliases where it still serves discovery. The prose carried the split’s history in a dozen places — a heading reading “Three channels, not one” against a note in which nothing was ever one, a paragraph explaining how the note was organized, “one qualification belongs here rather than in a later caveat,” and a question note saying the synthesis “now argues” its case. A note should read as an article about its subject, not as a record of how it came to be written. The same pass cleaned Scandinavian legal realism, which had picked up “two other notes carry the arguments this one declines to make.”

2026-07-31

  • Ran the queued F-Droid security-model research from Todo. All nine listed sources were read; the GrapheneOS tweet was recovered as Wayback JSON, the forum thread through its crawler view, and archive.ph/j7qql stayed unread but is characterized from both sides in the IronFox thread. Imported three artifacts: the two oss-security posts in fdroidserver pinning bypass disclosures (six demonstrated AllowedAPKSigningKeys bypasses, 2023–2025, root cause — custom certificate extraction diverging from apksig — unreworked through fdroidserver 2.4.4), wireguard inclusion policy violation (a policy-violating updater shipped six months unflagged, plus Jason Donenfeld’s “basically rotted” account of why WireGuard left F-Droid), and Malware in F-Droid build of Nextcloud News App (the one realized supply-chain incident: JitPack dependency poisoning that left source-to-binary correspondence intact, misattributed to F-Droid in the critique ecosystem).
  • The synthesis landed in F-Droid (new “Tooling and operational record” section: the pinning saga, the dependency incident, the leaked deploy key, scanning-not-review moderation in both directions, the AGP 8.12 buildserver lag and its November 2025 fix, and the specific shape of the developer defections), Android app distribution trust models (correspondence binds binary to source, not source to safety; the developer-signed path’s guarantee is only as strong as the bypassed pinning mechanism; “audited source” corrected to scanned-not-reviewed; a new open task tracks apksig adoption), PrivSec.dev on F-Droid security issues (the author’s January 2025 restatement, including his misattribution of the Nextcloud incident), and WireGuard (distribution-history backlink).
  • The net finding against the queued claim: “signs nearly every app” is down to about two thirds and falling; “outdated build infra” is a recurring-lag pattern whose named instance was retired in 2025; “poor moderation” is automated scanning rather than review, cutting in both directions; and no account, including GrapheneOS’s, alleges a compromise F-Droid itself shipped.
  • The user identified archive.ph/j7qql, the one source the pass could not read, as fdroid-website MR 834. It is the July 2022 permissions-blog discussion: the origin of the critique’s misleading-permission-display point (conceded in substance by the F-Droid author who opened the thread) and of the F-Droid–GrapheneOS feud, whose comment deletions each side remembers differently. Imported the rendered page and wrote fdroid-website mr 834, linked from F-Droid and PrivSec.dev on F-Droid security issues.
  • Ran the 2026 currency pass over the F-Droid critique the user asked for. Verified against current state: the stable full client targets API 30 under an in-code ExpiredTargetSdkVersion suppression (Basic targets 35, the 2.0 alpha rewrite targets 37); routine publish latency measured at two to five days (Catima, K-9 Mail, Organic Maps index-vs-upstream sampling), while the tail case is alive — Fennec and Mull carried 42 Firefox CVEs for two months in late 2024 under the toolchain purity rule; the website still lists raw install-time permissions (the MR 834 concession never shipped there); fdroidserver metadata is still single-signer, so key rotation is not first-class and adopting reproducible builds still costs a reinstall; the privileged-extension vulnerability claim remains publicly unsubstantiated and the mechanism is largely obsolete post-Android 12. Android app distribution trust models gained a “Currency check: July 2026” status table, F-Droid gained client-software and Fennec/Mull moderation paragraphs, and the PrivSec note’s durable-complaints line now points at the check.
  • Imported the BBC News report on Colin Gray’s sentencing (saved copy). Colin Gray is the father of Colt Gray, the 16-year-old who killed four people at Apalachee High School in Winder, Georgia in 2024. He was sentenced to 15 years — the third US parental criminal liability case for a child’s mass shooting, following the Michigan Crumbley prosecution. Wrote BBC Colin Gray sentencing report (source note) and Apalachee High School shooting (event note, seed), with a redlink for Parental criminal liability for school shootings.
  • Reframed the Apalachee note after recognizing that Judge Primm’s admonishment included failing to monitor the son’s online activity and failing to read his private notebook — two surveillance duties alongside the gun-storage failure. The case is not a pure firearm case with a theoretical digital parallel; the digital-monitoring obligation is already in the judge’s reasoning. Added a warning callout connecting the tension to the Parental control tools evidence review and evolving-autonomy literature.
  • Deep research on parental criminal liability and digital surveillance of teens. Imported three sources: Rethinking Youth Privacy (Citron & Waldman, Va. L. Rev. 2025), Behavioral Forensics parental liability white paper (July 2025), and Motivational framing of technology restrictions (Weinstein & Przybylski, 2019). Updated Parental control tools evidence review and Alternatives to age verification with the Weinstein/Przybylski concealment evidence and the Citron/Waldman structural critique. Wrote Parental criminal liability for school shootings as a concept note covering all three US prosecutions, the digital-monitoring extension in the Gray judge’s reasoning, and the unresolved tension between the gatekeeper-liability and evolving-autonomy positions. Added Setting Parental Controls Bernabe (source note) correcting the common misreading of a 2014 Georgia case.
  • Wrote Mandatory parental surveillance of adolescents as a synthesis note on the legislative wave making adolescent digital privacy structurally impossible through mandatory age verification, parental consent, and monitoring requirements.
  • Imported BBC News article Midnight social media curfew proposed for UK teens aged 16 and 17 (15 July 2026) reporting the UK government’s announcement of default midnight-to-6am social media curfews and off-by-default autoplay and infinite scroll for 16- and 17-year-olds, plus expert reactions from campaigners, academics, and the children’s commissioner. Renamed source artifact from a mistaken “signal-london-stabbing” filename.
  • Updated UK Online Safety Act to reflect the June 2026 under-16 ban announcement and the July 2026 curfew package, with expert reactions and BBC source citation.
  • Updated Australian under-16 social media ban and Do under-16 social-media bans reduce measured harm to reflect that the UK has moved from consultation to announced implementation.
  • Updated Minors and online regulation index to list the UK alongside Australia and France as a tracked national age-threshold regime.

2026-08-01

  • Ran a third design exploration for the wiki’s reading interface, this time blind to the interface it would replace. A brief stating only the content model, the corpus measurements, and the reading tasks went to four agents in clean contexts, each with a different starting question, and each returned one working self-contained HTML page. Imported all four as Wiki reading interface prototypes.
  • Moved the two earlier explorations into the vault as evidence, since they are what the new meta note argues from: Style-led wiki redesign mockups for the 2026-07-14 images and Feature-driven wiki mockups for the 2026-07-21 set. Their prompts are preserved verbatim under src/sources/; the overview and fit assessment that lived in share/mockups/*.md are now carried by the two source notes, and share/mockups/ is retired.
  • Wrote Reading interface redesign proposals, which records what the three explorations found and what was adopted. Its load-bearing finding is a measurement rather than a preference: the median note runs 342 words, 35 percent carry no heading at all, half carry two or fewer, and only 16 percent carry the six or more that an outline can serve. A layout tuned to the long synthesis is tuned to one note in fifteen.
  • Four designs that could not see each other’s work or the existing interface converged on four points: a flat list of 554 notes is not navigation, a note’s kind is a word rather than a colour, claim-level citations belong beside the claim while the sources inventory belongs at the note’s boundary, and the link graph is not a navigation surface. The third of these derives spatially the distinction Linking and citation conventions draws in prose, and no designer was told it mattered.
  • Adopted for implementation: figures and tables may exceed the prose measure, footnotes render as sidenotes, D’s visual system including its callouts, a collapsible left outline rail, and a landmark ribbon. The treatment of the note’s tail is still open, and the note records why putting provenance behind a tab would defeat the purpose the metadata serves.
  • Listed the meta note in Wiki architecture. The three source notes are reached from it rather than from the index, which lists destinations rather than evidence.
  • Replaced the adopted outline decision in Reading interface redesign proposals with D’s section spine and its running head, after the spine was seen working in a browser. Static screenshots had shown its geometry and none of its meaning, because it carries section names in title attributes, so it had been recorded as unlabelled gutter furniture. The spine costs about 28 pixels against a rail’s column, which settles the objection the heading distribution raised, and the same scroll pass drives the breadcrumb, so the two are one mechanism rather than two features. Recorded the label affordance and the heading depth it reads as the open questions that follow.

2026-08-02

  • Merged the two Business cluster diagrams in Graphs into one. They were split because a diagram could not render wider than the reading column, so the thesis-to-proposals path and the proposal branches were drawn separately. The renderer now sizes a diagram to the drawing it produces, bounded by the note rather than by the prose measure, so the cluster reads as the single tree it is: everything above the proposals hub establishes who is being sold to, and everything below it is a branch of what is sold. The prose no longer explains the split.
  • Updated Todo for the reading-interface redesign. Closed the redesign item against Reading interface redesign proposals, and opened four that the work left behind: verifying the layout below 900 pixels, verifying that a graph refits when its tab is first shown, covering the scroll wiring with a test, and deciding the header’s remaining differences from the prototype. The test item is the load-bearing one: three regressions during the redesign — a dropped progress function, a stale heading lookup, and grid placement that collapsed the article to 26 pixels — each passed the whole suite, because nothing asserts that the scroll handler is attached or that its selectors still resolve.

2026-08-03

  • Set a house standard for footnotes now that the renderer projects them into the margin as sidenotes, and swept the vault against it. A footnote is a bibliographic reference read without clicking, not a link with a label: it names who is speaking, what the work is, when, where inside it, and how to reach it, including the saved copy the wiki holds. Added the rule to AGENTS.md and to Linking and citation conventions, with the form the vault’s best notes already used — Realist foundations of Swedish rättssäkerhet deficits supplied the model.
  • Opened the margin to asides that are not citations: how a figure was counted, a gloss the sentence should not stop for, a divergent account and who holds it, a peripheral fact the sentence cannot carry. Two exclusions carry the risk. A caveat that fails the delete test is not rescued by demotion to smaller type, and anything a reader must not skip belongs in a callout, because a sidenote is skippable by construction and disappears below 1120 pixels. The line is about the reader, not the material.
  • Rewrote 97 thin footnotes across 48 notes. Every bare wikilink is gone: a footnote citing a source note now carries the work’s own identification beside the card, which is what a reader in the margin needs and what a note title alone cannot give.
  • Four defects surfaced while verifying citations, and were fixed. How Social Democrats legitimize cutbacks carried a DOI resolving to an unrelated article on Russian political trust, and described a Germany-and-Sweden comparison that the study does not make; the saved artifact holds the correct citation. Secret Service NTAC Protecting America’s Schools was dated 2020 from its upload path, against a report published November 2019. Swedish housing regime sources misspelled Peter G Håkansson and omitted the chapter title. In Unmasking pseudonymous public figures, the footnote on Del Naja’s denial cited a Yahoo article about Neil Buchanan that does not contain the quotation; the denial is reported by artnet News, and the Yahoo piece supports the Buchanan claim it now sits under. The Vega v. Tekoh link pointed at the wrong docket.
  • Footnotes to living project documentation now say so. Vendor and project pages state intended behaviour rather than audited behaviour, carry no snapshot under src/sources/, and can change between a note’s reviews; Cryptocurrency design comparison states this once in a callout and the rest name it per source.
  • Renamed Motivational framing of technology restrictions, whose stem transcribed the paper’s title at 84 characters — the vault’s longest source handle against a median of 37. Five MD013 reports pointed at it, all of them the same unbreakable wikilink, which is the linter reporting a filename rather than a wrapping fault. Four of its five call sites already piped the handle down to “Weinstein and Przybylski (2019)”, which is the sign a handle is doing the title’s work. The title is unchanged and the display text survived the rewrite.
  • Section wikilinks now render as Article § Section, and a same-page link as § Section, closing the Todo item that asked for both. A # is a URL fragment marker rather than a typographic one, and it set two names solid inside a sentence. An em dash and a parenthesis were both rejected: each carries sentence-level meaning, so mid-clause a reader parses it as punctuation of the sentence rather than as structure inside the link. The mark sits in a dimmed span, so the two names carry the label.
  • AGENTS.md now states where a section link can sit. Its label reads as two names joined, which behaves like a long proper noun: it belongs at the head of a sentence, after “See”, or ending a clause, and wedging one inside a noun phrase breaks the sentence carrying it. Length alone still does not earn a pipe.
  • Swept the six call sites that violated the new rule. Content-led privacy shop, Privacy venture roadmap twice, and Privacy legal and regulatory posture each wedged the PrivacyOps proposal into a noun phrase, and take the pipe the existing rule already allowed, since “PrivacyOps” is the name eight notes use in bare prose. Pornography literacy and harm reduction wedged two headings that are declarative clauses, and Minors and online regulation one; those three were recast so the link heads its sentence, which needed no pipe at all.
  • Added just link-subjects, which finds links whose target sits in the sentence’s subject slot. A link taking a verb only a document performs — shows, argues, develops, notes, records, sets out — has put the page where its subject belongs. The report separates targets by type, since a source note naming its own finding is citation prose and a concept doing the same is a category error, and it excludes meta and index notes, whose prose is about the wiki’s surfaces. It found 283 candidates across 175 notes, 135 of them naming a concept or entity.
  • Cleared the eight appositive cases the report isolated. A cross-reference appended as as <span class="unresolved">X</span> develops is grammatically identical to “as Lidén (2018) shows”, so the reader cannot tell cited evidence from the wiki citing itself. Each now states its claim and gives the cross-reference its own sentence, in Automated CSAM detection, Correction channel, Duress credentials and coercive extraction, Android Package, Children’s circumvention behaviours online, Encryption key disclosure compulsion, K-12 School Shooting Database, and Per Olof Ekelöf.
  • Correction channel now cites Lidén (2018) through the link rather than carrying the author’s name outside it, and the citation form is registered as an alias so search and the unlinked-mention detector both see it.
  • Split just link-subjects by verb class after the first sweep pass showed the raw count overstating the problem. A verb of exposition — explains, develops, describes, documents, treats, places — is something only a document does, so an abstraction taking one has been displaced by its page. A verb of evidence is not: two privacy companies charging money do show that the model sells, and every hit in Privacy competitors and inspirations was of that kind. The primary section now reports 98 rather than 135. A published work in the subject slot stays a false positive, since Adam och övergreppen documents what it documents.
  • Swept Rättssäkerhet in Swedish criminal cases, the note carrying the most hits. Five sentences now make their claim about the world and link the subject inside it, covering coercive interrogation, Kaj Linna, lay judges, duress credentials, and the right to silence. Two were left: remand detention does add a failure mode, and a book does document a case.
  • Swept a further 34 subject-slot links, taking the primary count from 98 to 64. The recurring shape was a signpost dressed as a claim: a concept named as the thing that covers, describes, or explains, where the sentence wanted the concept itself. Most resolved to a copula — therapy notes and biometrics are High-consequence data, the resulting doctrine is Swedish right to silence and förklaringsbörda — and the rest to putting the claim in the world and the link inside it. Touched Private AI strategy, Privacy business thesis, Data minimization, Why a known harsher penalty can beat an uncertain lesser one, Case for privacy and security, Privacy legal and regulatory posture, Dumpen, EU and Swedish data removal rights, Privacy product architecture, Why you shouldn’t talk to the police, and nine others.
  • Two exemptions found while sweeping, both correct as written. A section headed for neighbouring notes is prose about the wiki, so naming pages there is what it is for. A published work does document things, which is why Adam och övergreppen keeps its verb.
  • Swept a further 35 subject-slot links across 30 notes, taking the primary count from 64 to 29. The pattern held from the previous pass: a concept named as the page that covers or explains it, where the sentence wanted the concept. Local AI is the service boundary rather than explaining it, AzireVPN is the worked example rather than treating one, and pepper spray is excluded from Defence marker sprays rather than having its exclusion explained by them.
  • Two more exemption classes, both left as written. A “Reading path” section is navigation like the neighbouring-notes sections found earlier, so EU AI Act keeps its guide to Facial recognition and Gait recognition. A protocol separating two of its own properties is a design fact, which is why Ethereum still separates block inclusion from finality.
  • Finished the exposition-verb sweep, taking the primary count from 29 to 6 across 23 notes. What remains is the exemption set, and it is worth naming so a later pass stops re-reading it: a published work documents things, as Adam och övergreppen does; a navigation section names pages on purpose, as the reading path in EU AI Act and the neighbouring-notes section in Investigative confirmation bias do; a protocol separates its own properties, as Ethereum does; and an organization argues and describes in its own publications, as Obscura and PTS do.
  • One of the fixes from the previous pass was itself a hit. Analyzing Crime Displacement and Diffusion had been given its own sentence but still left the concept doing the stating; Displacement now rests on the argument rather than stating it. The same shape appeared in Pornography literacy and harm reduction, where an earlier repair traded one exposition verb for another.
  • Ran a targeted pass over the weaker-signal half of just link-subjects, which a census showed was mostly sound. A comparison note does compare, place, treat, and identify, so the 106 exposition verbs against synthesis targets were largely honest cross-references and stayed. What did not survive was explains, where the sentence carries no content of its own: “EU and Swedish data removal rights explains these limits” tells a reader only that an explanation exists somewhere. Twenty-one notes changed; the total fell from 176 to 155.
  • The distinction the census exposed is informativeness, not grammar. “compares XMPP with Matrix and dedicated private messengers” names what sits at the other end of the link; “explains these limits” does not. No diagnostic can find that, so the remaining flags are the tool being imprecise rather than the prose being wrong, and sweeping them would trade honest cross-references for contortions.
  • Renamed Should you talk to the police to its present handle, with the subtitle “how asymmetry favours silence”, and inverted the opening so the claim leads. The old handle was interrogative on a synthesis note and collided with the Should innocent people talk to the police source card beside it. The opening now asserts the recommendation and its scope — when you are, or could become, a suspect — and makes the critique of Duane’s evidence the argument rather than a qualifier on it: the harm is documented and severe, the benefit is unmeasured, and the law punishes an ambiguous silence. The recommendation survives that critique; the argument usually made for it does not.
  • Narrowed the subtitle rule in AGENTS.md, which had overshot its own reason. It was written to stop just new and the docs modelling a subtitle for every note, and it forbade one outside source notes on the ground that a synthesis handle is already a claim. That holds for a restatement and not for a warrant, so the field is now legitimate anywhere it carries what the handle cannot. This is the vault’s first non-source subtitle.
  • just rename rewrites wikilinks but not Mermaid node labels or %% note-link markers, so Graphs kept a stale label and directive after the rename. Both are corrected, and the rendered page was checked in a browser: nine fences, nine SVGs, no syntax error, and the node label and link both resolve to the new page.

2026-08-04

  • Imported and fact-checked Ken LaCorte’s video “The most avoided number in America” (transcript saved as src/sources/2026-08-04-lacorte-most-avoided-number-transcript.txt), verifying its claims on deterrence, policing, wrongful conviction, and father absence against the scholarly literature. The source card The most avoided number in America records the verdicts: victimization statistics accurate (BJS 2023, del Pozo 2022), NYC figure right but sourced to the wrong report, the 2-percent-of-counties claim advocacy-sourced, and the welfare-incentive narration contradicted by the video’s own source document.
  • Created four concept notes from the verification pass: Deterrence (certainty beats severity; NRC 2012 finds the death-penalty literature uninformative; the effective margin is enforcement, not sentencing), Homicide clearance in the United States (six-decade decline, the racial gap carrying it, surveillance tech evaluations null — review_after: 2027-08-04 set for the clearance figures), Father absence (Sampson’s mediation mechanism, modest causal effects after selection controls, the AFDC-abolition timeline refuting the welfare-incentive claim), and Wrongful conviction (Gross et al.’s 4.1% death-row floor, the NRE race report’s exoneree distribution).
  • Opened the question note What explains the 2020 US homicide spike, keeping the video’s settled police-pullback attribution open: Kegler 2022 and López/Rosenfeld constrain the timing, Devi & Fryer is a contested working paper, and no design yet separates pandemic shock, pullback, and legitimacy channels.
  • Cross-linked What if boys were never the problem (same creator) and Correction channel (exoneration-registry selection).
  • Saved four verified PDFs under src/sources/ (Sampson 1987, Nagin 2013, Gross 2014, Chalfin 2022). Three sources could not be downloaded (NRC 2012, Cook & Mancik 2023, NRE race report 2022: 403/anti-bot blocks); they are cited by DOI/URL with the limitation recorded in the notes’ footnotes. One planned citation (Pyrooz et al. 2016 “Ferguson effect”) failed verification in OpenAlex and was dropped.

2026-08-06

2026-08-08

  • Deep-researched the comparative question “does Sweden’s rättssäkerhet stand out internationally, or are hidden miscarriage flaws ubiquitous among high-ranked democracies?” New synthesis: Hidden miscarriage risk and correction-channel opacity across democracies. Verdict: the epistemic structure is ubiquitous (no country can count its wrongful convictions; visible corrections measure the correction channel, not the error rate; WJP-style rankings are perception instruments with ~29 expert respondents per country and no variable that could register a hidden miscarriage), but Sweden sits at the opaque extreme of its Nordic peer group — the only one with no independent or special reopening body, no published channel statistics, and corrections driven entirely by journalism. Norway’s commission (433 reopenings 2004-2021) versus Sweden’s 40-60 resning grants per decade is the natural experiment; Sweden’s ~98% prosecutor versus ~1.25-2.5% unrepresented-private grant asymmetry is the sharpest datum.
  • Imported eight sources with cards: Resning i brottmål i Sverige (Lidén, SvJT 2021; Swedish grant-rate table 1935-2020), Exoneration in Sweden (Martinsson, ELR 2021; the transparency deficit), Norwegian Criminal Cases Review Commission annual report 2021, Between Legal Certainty and Doubt (Dutch novum reform five-year evaluation), Prevention of Wrongful Convictions (Stridbeck & Magnussen on Norwegian safeguards), WJP Rule of Law Index 2025 methodology and Sweden profile, Japan’s Hostage Justice System (HRW 2023), and National Registry of Exonerations 2024 annual report (URL-only; the registry blocked all download routes, same as the 2022 race report).
  • New entity notes: Gjenopptakelseskommisjonen (Norway’s commission), Criminal Cases Review Commission (UK CCRC, including the “vanishing” referral rate, the Exeter weak-evidence finding, and the Horizon scaling failure), and European Registry of Exonerations (EUREX).
  • New event notes: Baneheia case (Kristiansen acquitted 2022 after six rejected applications; the commission model’s own limits) and Guðmundur and Geirfinnur case (Iceland’s solitary-confinement confessions; five acquitted 2018, 44 years on).
  • Updated: Correction channel (comparative field-test section), Resning in Swedish criminal cases (grant asymmetry, commission links), Rättssäkerhet in Swedish criminal cases (international-frame section closing off both “Swedish pathology” and “therefore ordinary” misreadings), Wrongful conviction (registries beyond the US), and Coercive interrogation and false confessions (Iceland case added as fifth documented system).
  • Redlink left for Post Office Horizon scandal.
  • Open research queued in the synthesis: Denmark’s Klageret grant statistics, EUREX per-country dataset (script-rendered, unfetchable), Finnish reopening volumes, Canada’s new commission statistics, and the SvJT 2021/574 Norwegian-law-history import.
  • Deep-researched the comparative privacy question “how unique is Sweden’s public-by-default identity system (offentlighetsprincipen + personnummer + utgivningsbevis) internationally — is Sweden systematically doxxing every citizen irreversibly?” New synthesis: Public-by-default identity in international comparison. Verdict: every component has a close foreign analogue (universal register identifiers are Nordic-normal; Norway and Finland publish tax data; Germany sells address lookups; Iceland ran a fully open registry; US brokers dwarf Swedish people-search), but the composition is singular — Sweden alone combines a misuse-unresettable universal identifier, a state distribution backbone (SPAR), anonymous unlogged access, retail conviction search, and paid subject-monitoring alerts. Sharpest new datum: Sweden is an outlier even among exposed systems on reversibility — Denmark (identity-misuse reset), the US (SSN reset for abuse victims), and Japan (My Number leak reset) all offer the escape Sweden refuses.
  • New concept note Access-transparency: subject-visible query logging as privacy architecture (Norway’s tax-list snokevarsel cutting searches ~85%, Estonia’s Data Tracker going mandatory, Sweden’s own 1973 credit-report copy that was never generalized).
  • Imported twelve sources with cards: Search the tax lists (Norway’s watched-search model), Public disclosure of tax information (SSB DP 975; the ~85% search collapse), When can you process personal data (Danish CPR confidentiality), Public information on individual income taxes (Finland’s terminal-fenced tax openness), Estonian Data Tracker (RIA page via pure.md after origin 403), Can I change my Social Security number (SSA via Wayback after origin 403), The electoral register and the open register, Bundesmeldegesetz 44 simple register extract, We Know Your Identity Number (Grapevine 2005; the openness-as-security counter-case), My Number FAQ (Japan’s locked identifier + leak reset), Kan jeg få et nyt personnummer (Danish misuse reset), and Folkbokföringens historia (Skatteverket’s 1947 origin account).
  • Updated: Swedish public-record privacy (international-comparison section), Swedish personal identity number (the misuse-reset gap), and Offentlighetsprincipen and personal privacy (practical obscurity’s US origin + comparative pointer).
  • No new redlinks; open research queued in the synthesis (Icelandic and Finnish reset rules, Norwegian fnr resets, dossier-industry size comparison, C-199/24’s reach beyond conviction data, protected-identity scheme comparison).
  • Corrected the comparison’s disconfirmation section after user challenge on three points where it was too generous to Sweden: (1) the state does not merely “not dump” personnummer — it legally compels registration at the true sleeping address (folkbokföringslagen 7 §, 25 §, 42 §; NJA 2022 s. 287 upholds a conviction) and sells the lookup through SPAR; (2) health-data secrecy has a statutory court channel — the RPU act (1991:1137, 14-14 a §§) overrides medical and social secrecy and the evaluation enters the public court file; (3) the FUP, not the mugshot, is the exposure — Åklagarmyndigheten confirms the investigation file becomes public at charge, and Domstolsverket reports far more than 100,000 FUPs released yearly after per-request masking review (its AI-masking news item), with victim-oriented secrecy leaving the accused’s own intimate material reachable. Imported both FUP sources; all three corrections verified and written into Public-by-default identity in international comparison, whose “composition is singular” finding now includes the criminal-enforced address duty.
  • Restructured the comparison’s calibration section after user review: “Where the doxxing charge overshoots” no longer matched its contents (the corrected bullets strengthen the charge), so it is now “Testing the doxxing charge in both directions” with “Sharper than the charge” and “Where the charge overshoots” subsections. Fixed footnote wikilinks I had written across line breaks (invalid per house rules) in Public-by-default identity in international comparison, Access-transparency, Gjenopptakelseskommisjonen, and Hidden miscarriage risk and correction-channel opacity across democracies, plus one pre-existing instance in Privacy gear catalog; a broken wikilink inside sources/2026-08-01-wiki-reading-interface-design-brief.md stands because sources are immutable.
  • Added the user’s epistemic-authority argument, which survives verification: US broker dossiers are third-party claims (ad-tech inference and marketing fragments around official pieces, per the FTC source) that no institution warrants, giving the American subject plausible deniability; the Swedish dossier reproduces legally compelled, state-certified records — trusted when weaponized and authoritative until the state itself corrects them. Folded into the “US is worse on some axes” calibration (now “worse on some axes, and weaker in authority”), the composition finding, and the comparison table’s dossier cells.

2026-08-09

  • Imported Why not matrix, haru’s twenty-one-item critique of the Matrix protocol, from nonogra.ph (rendered and raw snapshots under sources/2026-08-09-why-not-matrix.*). The import’s decisive finding is provenance: the page carries a 2026-07-23 byline but its prose is byte-identical to an article haru published on Telegraph on 7 August 2023, discussed on Lobsters that September in a thread that quotes item 11 verbatim and drew a reply from Matrix’s project lead. Nothing on the republished page marks it as archival.
  • Verified the article’s claims against the specification and release notes rather than accepting them. Three families are overtaken: media downloads stopped being unauthenticated by default when MSC3916 shipped in Matrix 1.11 (June 2024) and Synapse 1.120 flipped the config (November 2024); state resets acquired the fix the article says does not exist when State Resolution v2.1 shipped in room version 12 (September 2025); and administrators losing power to resets was closed in the same room version by giving creators an unchangeable infinite power level. The canonical-JSON item was wrong when written — the spec has defined Canonical JSON, with a grammar and a reference implementation, since November 2021. The permanence group survives intact, because it describes the data model rather than an algorithm.
  • Rewrote Matrix around the append-only event graph, which the note previously did not mention at all despite it driving every property the note did discuss. Added state resolution and state resets, room version 12 and Project Hydra with its own stated limits, the media history, and CVE references.
  • Extracted Advisory deletion as a concept. Eight notes were restating the same premise in vocabulary that barely overlaps — Nostr, ActivityPub, AT Protocol, Secure Scuttlebutt, Redact.dev, DeleteMe, Incogni, and decision rule 5 of Secure and decentralized communication stacks — and none linked a common page because none existed. The note is written at the general scope (protocols, platforms, and brokers alike), grades deletion on five rungs from enforced to impossible, and records the counterintuitive part: a retraction must propagate as widely and durably as what it retracts, so the tombstone routinely outlives the thing it buries.
  • Connected the restatements to it rather than leaving the page orphaned, and paired it with EU and Swedish data removal rights, where GDPR’s duty to take reasonable steps to inform other controllers is the same structure written as a legal obligation.
  • Left Crypto-shredding as a redlink. It is the one mechanism that defeats advisory deletion, and this session named it without grounding it.
  • Not verified: the article’s item 10, that device-list updates leak client and platform information. It is load-bearing for nothing the vault now asserts, so it was left unchecked rather than repeated.

2026-08-10

  • Wrote the three source cards the 2026-08-08 import left empty. Can I change my Social Security number, Estonian Data Tracker, and The electoral register and the open register existed as frontmatter and nothing else: just new created them, that session’s log recorded twelve sources imported “with cards”, and three of the twelve were never written. Their evidence was already saved under src/sources/ and already cited from Access-transparency, Public-by-default identity in international comparison, and Kan jeg få et nyt personnummer, so for two days the vault cited pages with nothing on them.
  • Each card describes its artifact and then what the artifact cannot settle. The Estonian page is the Information System Authority’s account of a service it operates, so it establishes the architecture and not its coverage — and coverage is the load-bearing question, since a tracker some registers implement and others do not tells a subject nothing about the ones that stayed out. The claims the wiki attaches to the tracker beyond that page, its 2017 arrival on eesti.ee, the § 36 logging duty in the Estonian Personal Data Protection Act, and the 2025-26 mandatory-adoption bill, rest on the sources cited alongside it, which the card now says. The Social Security Administration’s answer lists five grounds for a replacement number and nothing about thresholds, timelines, or whether a reset severs the old number’s history in records the agency does not control. The GOV.UK guide gives the open register’s opt-out and the full register’s five permitted uses, and is silent on enforcement and on copies already sold.
  • Each title now carries its work’s own, with the derived handle kept as an alias: the agency page prints a question mark a filename cannot, the GOV.UK guide quotes ‘open register’, and the Estonian page is titled “Data tracker”. Nothing changed in the notes that cite them; their footnotes already pointed at the saved artifacts.

This note cites no sources of its own.

Working out connections…