Field Notes

concept

Correction channel

StatusStableUpdatedStands on6 sourcesSupportsnotes

A correction channel is the set of processes by which an institution’s errors come to light and get remedied: appeals and reopening mechanisms, internal review, regulator complaint procedures, platform appeals, and — in systems whose internal routes are weak — external investigation by journalists and advocates. Every institution that produces judgments has one, and its output is visible: a count of errors discovered and corrected.

The concept matters because of what that count measures. A correction channel’s throughput depends on its own properties — the threshold for opening a review, the resources available to pursue one, who is entitled to ask, and how much external pressure the institution faces — not on how many errors the institution actually makes. The count of corrected errors is therefore a measurement of the channel, not of the error rate.

Discovered errors are a discovery count

Three readings of a correction record are common, and two of them are unsound.

  • “Few corrections means few errors.” A low correction count is equally consistent with a low error rate and with a blocked, narrow, or discouraged channel. Where reopening requires the convicted person to produce new material before institutions committed to the original outcome will act, as Sweden’s resning structure does, the threshold itself suppresses the count.
  • “Many corrections means many errors.” A high count is equally consistent with a high error rate and with an unusually active channel. An exoneration registry observes only the errors the correction system succeeded in recognizing, and growth in exonerations can reflect better discovery rather than worse adjudication.
  • The defensible reading is about the channel itself: corrections document that the channel operated in those cases, at that threshold, under that pressure. They describe the correction mechanism’s throughput, and they bound nothing about the errors that never entered it.

The informative set is often the one the channel refused. A denied application backed by recanting witnesses marks where the channel’s threshold sits, which is why Rättssäkerhet in Swedish criminal cases reads the denied resning in Billy Butt as evidence about the Swedish channel’s operating point rather than about anyone’s guilt.

An institution cannot calibrate itself

The deeper problem is circularity. Validating an institution’s accuracy requires comparing its outputs with independently known outcomes. When the only available labels are the institution’s own — convictions counted as verified true accusations, acquittals counted as verified false ones, final judgments treated as factual ground truth — the institution generating the labels is the institution whose reliability is in question. Finality establishes legal responsibility; it does not supply the independent ground truth that calibration requires.

The archive study in Lidén (2018) is the vault’s direct warrant for this point. Her system-level estimates imply that many wrongfully convicted people who appeal or seek a new trial will not be acquitted under a wide range of assumptions, which directly challenges the use of successful exonerations as the denominator for system error. The two JK commissions reached the matching conclusion from the case-file side: the 2006 project examined a selected set of known failures and could reveal how errors survived, not how often they occur; the 2009 report’s small later samples did not establish system-wide failure rates either.

Samples drawn from a channel are selected on it

Any sample assembled from a channel’s outputs inherits the channel’s selection. Statistics computed over exoneration samples — the share of exonerees who falsely confessed, the share who were juveniles — describe risk factors and relative frequencies among recognized errors, not the frequency of error among all cases. Gross, Jacoby, Matheson, Montgomery, and Patil’s 340-case study of US exonerations from 1989 through 2003 is the canonical example: its figures are real, but the sample is everything the channel caught, not a draw from all convictions.1

The same structure appears wherever evidence is assembled by the people it later vindicates. A defense attorney’s archive of clients who talked to police and were burned by it excludes by construction everyone who talked, was believed, and was never charged, which caps what the archive can prove. Why you shouldn’t talk to the police works through that limit. Police classifications of accusations as “confirmed false” count only the accusations investigators could and chose to classify that way. False accusations as an epistemic risk works through what that leaves unmeasured.

The comparative field test

Channel architecture varies by an order of magnitude across countries whose true error rates cannot be compared, and the variation tracks institutions, not case quality. Norway’s independent Gjenopptakelseskommisjonen reopens 15-18 percent of merits-reviewed applications and publishes everything annually; Sweden’s court-internal resning granted about 1.25 percent of unrepresented factual claims at Högsta domstolen in 2005-2010 and publishes nothing. The United States built a detection industry — a registry, innocence organizations, prosecutorial review units — and its exoneration count grew exactly as that industry grew. The four architectures and their national variants are compared in Hidden miscarriage risk and correction-channel opacity across democracies.

What would measure error instead

An error rate needs a denominator the channel did not select: independent ground truth applied to a defined reference class. Examples of the shape required:

  • retesting a random sample of closed cases against evidence the original process did not use, as DNA testing made possible for a fraction of convictions;
  • audit of cases against independently auditable traces whose provenance and error processes can themselves be tested;
  • calibration studies comparing expressed confidence with known outcomes across comparable cases.

Where none of these exists, the honest product is not a guessed prevalence number but an analysis of failure mechanisms, exposure, detectability, and severity — the posture False accusations as an epistemic risk adopts for exactly this reason.

  1. Samuel R. Gross, Kristen Jacoby, Daniel J. Matheson, Nicholas Montgomery, and Sujata Patil, “Exonerations in the United States 1989 Through 2003,” Journal of Criminal Law and Criminology 95, no. 2 (2005): 523-560, https://doi.org/10.1177/0022427804270584. ↩

Built on 6 sources (5 archived here, 1 external).

Working out connections…