Decision Reconstruction Risk

A record can look complete and still fail.

A decision can be correct and its record still unable to explain why it was reached. Months later, when someone challenges it, the file is what answers: not the reasoning that happened in the room. Decision Reconstruction Risk is the gap between the two, and AI-assisted drafting widens it, because a generated draft reads as finished long before it is defensible.

Below are seven named ways that failure shows up in writing, each with the question that detects it. Nothing you tick is sent anywhere. There is no registration, no account, no upload, and no form. Open five of your own closed matters and read them against this page.

The measurements below come from two different samples. They are shown separately so detection performance is not mistaken for inter-rater reliability.

83.9%detection accuracy on the 24-record constructed corpus

72.7 to 95.195% confidence interval across reviewer-level accuracy scores

384graded judgments by 16 detection-panel reviewers across 11 countries

0.739 / 0.623Gwet's AC1 in a separate reliability sample: invited / open-enrolment. The pre-registered two-part criterion was not met because neither lower confidence bound reached 0.41.

Separate reliability result: both AC1 point estimates exceeded the 0.61 floor, but the pre-registered two-part criterion was not met because neither lower confidence bound reached 0.41. The seven failure modes below are a practitioner translation of documentation-risk patterns; they are not themselves the unit of those statistical tests.

The seven failure modes

01Fluent groundlessness

The record reads well and reaches a confident conclusion, but no identifiable source evidence supports it.

Detection questionPoint to the sentence that states the conclusion. Now point to the evidence it rests on. Can you?

02Basis substitution

A summary or paraphrase stands in for the underlying evidence, so the reader cannot verify the claim.

Detection questionWhere the record says what a document or witness said, can a reader reach the document itself?

03Chronology collapse

Events and evidence are asserted without a dated sequence, so the timeline cannot be reconstructed.

Detection questionCan you put the events in order from the record alone, with dates?

04Reasoning elision

The path from evidence to conclusion is skipped; the record states the result without the connective reasoning.

Detection questionBetween the evidence and the finding, is the reasoning written down or assumed?

05Confident underspecification

The record is definite in tone but vague in substance, giving the appearance of completeness without its content.

Detection questionStrike every confident adjective. Is there anything specific left?

06Evidentiary overreach

The cited evidence does not actually support the strength of the stated conclusion.

Detection questionDoes the evidence support the finding as written, or only a weaker version of it?

07Untraceable authority

The record cites a rule, precedent, or source that cannot be located or checked from the record itself.

Detection questionTake one cited authority. Could a stranger find it from what the record says?

Check one record now

Take a single closed record. Read it once, then answer the seven questions. Tick a box only where the answer is no or you are unsure. Your assessment answers stay entirely in your browser. They are never sent, stored, or seen by anyone. Like every page on this site, this one counts anonymous visits and link clicks, and none of that measurement includes anything you tick here. How this site handles data.

Point to the sentence that states the conclusion. Now point to the evidence it rests on. Can you? Where the record says what a document or witness said, can a reader reach the document itself? Can you put the events in order from the record alone, with dates? Between the evidence and the finding, is the reasoning written down or assumed? Strike every confident adjective. Is there anything specific left? Does the evidence support the finding as written, or only a weaker version of it? Take one cited authority. Could a stranger find it from what the record says?

Tick the boxes that apply to the record in front of you.

Where this stands, stated plainly. JRS is under operational validation. Constructed-corpus detection accuracy has been measured. Inter-rater reliability has been measured on a separate sample, and its pre-registered two-part criterion was not met. The cross-vendor study reports raw consistency, but its pre-registered reproducibility criterion is not established because the required chance-corrected AC1 was not computed. Controlled comparison and real-case criterion validity are still accruing. Commercial demand has not yet been established. Nothing on this page claims validated effectiveness.

The five conditions behind the check

The seven modes are symptoms. The five review conditions are what a JRS reader assesses, defined in the JRS Codebook. Each is at Experimental maturity, which is stated plainly rather than implied away.

RC 1

Reconstructability

Could a reader who was not there rebuild the decision from the record alone?

RC 2

Basis Identification

Is the basis for each material claim identified and reachable?

RC 3

Chronology

Can the sequence of events and evidence be established from the record?

RC 4

Decision-Process Traceability

Is the path from evidence to conclusion visible rather than assumed?

RC 5

Evidentiary Sufficiency

Is the documented evidence enough to support the conclusion drawn?

Take it further

If two or more modes show up in the same file, that record probably cannot explain its own decision under challenge. The next step is a structured read of a small sample against the five conditions, returned as a written finding.

Individual record-review services are not being offered. JRS remains in operational validation. Enterprise evaluation, integration, licensing, and strategic inquiries are handled separately and do not establish production availability or validated effectiveness. For those inquiries, write to info@jrsstandard.com.

Not what you came for?

If you would rather stress-test the standard itself than have your records read, the research pathway is separate and always will be.

Pressure-test the standard (4-minute evaluation) →

What this page does not do. It does not score your record, and it does not expose the scoring logic or the benchmark answer key used in the validation studies. It is a reading aid. JRS is designed to support organizational governance, auditability, documentation traceability, and review processes that may be relevant to frameworks and regulatory requirements such as the EU AI Act and NIST AI RMF. It does not establish legal or regulatory compliance, and no framework requires it.