The vocabulary

Defined terms

Twenty-two terms, alphabetized. Each anchor is linkable directly.

Admissibility #

The organizing standard of the whole argument, borrowed from evidence law and applied to consequential decisions made at scale. Admissibility is not truth; it is the prior discipline a claim must satisfy before its truth can be fairly tested — that it was gathered, recorded, and preserved so that it can be brought into a forum of contest and examined by someone who declines to take the maker's word. A claim can be true and inadmissible, or admissible and false. See alsoThe receipt and the rebuttal

Audit log vs. account #

An audit log is written by the system, for its operators, to record that events occurred during normal running. An account is written for a reader who does not trust the system: it preserves the actual inputs as they stood, lets the decision be re-run, and answers to the person with standing to contest it. The materials look similar; the specification is different, and the specification is the point. See alsoAudit logs are not accountability

Authorization vs. legitimacy #

Authorization is permission to make a decision — the power, conferred by role or contract or statute, to act. Legitimacy is the standing to have that decision accepted as fair by the person it lands on. The two come apart routinely: a system can be fully authorized to impose a consequence and still owe an account of why it was right. An authorized decision answerable to no one is power without legitimacy. See alsoAuthorization is not legitimacy

Calibration vs. confidence #

Confidence is how sure a system sounds; calibration is how often it is right when it sounds that sure — the correspondence, measured over many cases, between stated certainty and actual accuracy. Machines have made confidence free and detached it from cost, so fluent assurance no longer carries the information it once did. An uncalibrated system can be perfectly confident and routinely wrong, and the confidence will not warn you. See alsoConfidence is not calibration

Chain of custody #

The companion to provenance: the unbroken, recorded account of what happened to a piece of evidence between its origin and its use, so the thing examined in the forum is demonstrably the thing that was gathered, neither swapped nor silently altered. Provenance establishes where a claim's inputs came from; custody establishes that they arrived intact. A break in custody is a hole through which the connection between a claim and its origin can quietly drain. See alsoThe chain of custody we never built for software

Contestability #

The property of a decision that lets it be argued against on its merits rather than merely complained about. A decision is contestable when the materials needed to challenge it — the inputs, the governing rule, the reproducible process — have been preserved and made reachable by the challenger, so the person who would dispute it can press on something solid rather than shout at something opaque. It is what provenance, replay, and standing amount to seen from the wrong end of the decision. See alsoA right you cannot exercise

Decision Receipt #

The record a consequential decision leaves behind, built to be contested rather than merely believed. It carries the real evidence the decision rested on — in the order it was consulted — the rules that were active at the time, and enough state to replay the decision and check whether the outcome follows. A receipt that carries those things is more authoritative than any memo, precisely because it invites the attack and withstands it. See alsoThe receipt and the rebuttal

Defaults-as-decisions #

The recognition that a standard for machine-made authority is being set right now by accumulated practice rather than deliberate choice, and that declining to choose is itself a choice. When an institution automates a consequential decision and preserves nothing examinable, it establishes a standard — that the model said so suffices. Each such act is a precedent, and the precedents harden, decision by decision, into an expectation that becomes first normal, then required. See alsoThe decision nobody made

Drift #

The silent movement of inputs, rules, or models away from the state in which a decision was made. Data is updated, a policy is revised, a model is retrained — and a record reassembled afterward confronts a different decision than the one that was actually made. Drift is a silent edit: it changes the basis of a past decision while changing nothing about the record that purports to describe it. The remedy is to freeze the world the decision was made in. See alsoDrift is a silent edit

The four-question test #

The four properties an admissible decision must carry, posed as questions a record must answer. Where did the inputs come from, frozen as they stood — provenance. Can the decision be re-run against them — replay. What rules governed it at the time — the rules captured in the form they held. And who has the entitlement to demand the record and a forum to contest it — standing. A record that cannot answer all four is incomplete. See alsoThe receipt and the rebuttal

The half-life of a justification #

A reason for a decision is not timeless. The conditions that warranted a conclusion decay: evidence ages, rules are revised, a model is retrained, relevance erodes. A justification that was sound when made can quietly become unsound as the world beneath it moves on, and it expires at the speed of its fastest-moving part unless the institution deliberately freezes the world the decision was made in. Admissibility is held against time, or it is lost. See alsoWhat a timestamp promises

Hearsay (applied to model output) #

In the law of evidence, hearsay is a statement offered for its truth that the forum cannot test by examining the one who made it. Model output has the same structure: it asserts, often fluently and with the production values of confidence, but there is no maker present to cross-examine and no standing on the line. The model said so is hearsay dressed as testimony. The remedy is not to ban the output but to refuse to credit it on its own say-so. See alsoHearsay at scale

Hindsight contamination #

The distortion that enters when a decision is reconstructed in light of its outcome. Once the result is known, the record tends to be read — and often rewritten — as though that result were always visible, and the decision is judged against facts that surfaced only afterward. A decision built for faithful reconstruction speaks with the authority of a witness who testified before the verdict was known; a decision merely logged enters the examination already shaped by what has been learned since. See alsoReconstruction beats recollection

The information horizon (what was knowable) #

The boundary of what could be known at the moment a decision was made. A decision can only be judged fairly against the evidence available inside that horizon — not against facts that arrived later, however damning they look in retrospect. Preserving the horizon is what makes a fair reconstruction possible: it fixes which inputs were actually reachable, so the decision is examined as the thing it was rather than as a present-tense account of a world that had not yet revealed itself. See alsoReconstruction beats recollection

Jurisdiction of a number #

The bounded set of decisions a score or model output was built and validated to govern. A number has authority only within the question it was designed and tested to answer; carried outside that jurisdiction — a risk score repurposed for a use it was never validated against — it lends false weight to a decision it was never warranted to make. Asking what a number is for, and refusing it where it does not belong, is part of taking it seriously. See alsoThe map is not the decision

Legibility #

The property of a record that a second reader can actually follow. A record can be complete and still illegible — its inputs buried in operational noise, its rules implied rather than stated, its path impossible to retrace. Legibility is the discipline of making the inputs, the governing rule, and the route between them plain enough to be reconstructed by someone with no access to the maker's intentions. A record no one can read is, for the purposes of contest, a record that does not exist. See alsoWho reads the receipt

Provenance #

The demand that a claim arrive attached to its origin, so what it rests on can be traced back to where it came from rather than summarized after the fact. The real sources are preserved in the order they were consulted and frozen as they stood at the moment of use, because inputs drift and order matters — a fact read before a contradicting fact does different work than the same fact read after. Provenance is treated not as a logistical nicety but as a moral one. See alsoProvenance is a moral category

Reconstruction #

The disciplined rebuilding, by someone who was not present, of what a decision was and why — and the commitment to build decisions so that this rebuilding can be faithful rather than filled with hindsight. Reconstruction is the future tense of admissibility. The aim is not to preserve a description of the decision but to preserve the world the decision was made in, so that when it is questioned — often years later, by people who never met the decider — it can still be seen as it was. See alsoReconstruction beats recollection

Remedy #

The means by which a successful challenge actually changes the outcome. Standing supplies the right to contest and a forum to be heard; remedy is what the forum can do when the challenge succeeds — reverse the decision, correct the record, repair the harm. Without remedy, a right to contest is a complaint with nowhere to land. A record built to be challenged is only as real as the correction its challenge can compel. See alsoReversibility is the other half of accountability

Replay (replay-determinism) #

The demand that a decision be reproducible against the conditions under which it was actually made — the frozen inputs, the frozen rules, the preserved state — so it can be run again and watched rather than merely described. A story asks to be believed; a replay asks only to be watched. Replay-determinism is the property that makes this trustworthy: the same inputs run back through the same process land where they landed before. Determinism is not discovered in a system; it is a constraint imposed on the decisions that need it. See alsoReplays, not explanations

The second reader #

The person, not present at the decision and not party to its making, who later opens the record and tries to understand what was done — the auditor, the challenger, the reconstructor, the affected party seeking the basis of a consequence imposed on them. The second reader is a stranger with reason to doubt, no obligation to believe, and no access to the maker's intentions except through what the record preserves. The working test of any decision is whether it would meet this reader as the thing it actually was. See alsoWho reads the receipt

Standing #

The entitlement of the affected party to demand the record, a forum in which to contest the outcome, and a remedy to make the challenge worth bringing. A record is only half of accountability; the other half is a contestant with the means to actually challenge it. A record answerable to no one is not accountability — which is why standing cuts the other way too, against a record open to everyone. It is built for the person with a reason to look. See alsoStanding: who gets to contest a machine

— Dispatches · Summit Cognitive