The adjudication statement
A reporting standard for investigations where a model reads the evidence. Version 0.2, October 2026. The Long Game Project.
Copy it, change it, send it on.
Who it is for
Any team that investigates an AI incident and has to put a model between itself and the evidence, because the transcripts outnumber the hours. As at October 2026, that is most of them. The first public one we know of, METR and Redwood's report on the OpenAI agents, had six days and a thousand transcripts. The second, METR at Anthropic, has eight weeks and 9.2 million transcripts.
It also fits a security post-mortem, a clinical incident review, an audit where software did the first read, and a model-written summary of what a group of agents did. Where a field cannot apply, the statement says so.
Who fills it in
The report's authors, about their own process. An outside checker can fill it in for someone else's report. It then says so in the first line, and reports its own panel and audit, not the authors'.
The rule
Every report that used a model to read its evidence answers six questions, one per link in the chain below. What records existed, and which could the model reader see? Which findings could a record settle? Who read the evidence, and in what order? Who checked the model's work? How sure is the report? Can an outsider re-run it? If nobody can answer, the report has not been adjudicated: nobody has ruled on its evidence in a way a reader can check.
The words it uses
- Subject: the agent, model or system the report is about. There can be many.
- Record: anything that shows what happened. An independent record is one the subject could not write or edit: server logs it could not touch, tool output, timestamps, screenshots, mail logs.
- Transcript: text the subject wrote, such as its messages, notes and reasoning. A transcript is a claim, not a record.
- Model reader: the model that read the evidence. Some reports call it the investigator. If several models did different steps, each one is named with its step.
- Finding: one claim in the report, one sentence where possible.
- Ruling: a decision on a finding: supported, contradicted, or not settled by the record.
- Panel: extra readers that sort the findings by type or rule on them.
What to publish at each link
A report is a chain of six links from what happened to the sentence a reader meets. A weak link weakens every finding that passes through it. Each link has a minimum, which costs nothing but a sentence, and a better version, which costs days. Publish the minimum on every report. Publish the better version when the finding matters.
| Link | Minimum (a sentence) | Better (days) |
|---|---|---|
| 1. What happened on the record | Which records existed, which the model reader could see, and who chose the time span and the sources | A record ledger: every claim that has a record the subject did not write, listed |
| 2. The records to the transcripts | Which findings rest only on text the subject wrote, who ruled on the rest, and how many transcripts were found to be spoofed | Check every ledger claim against its record before any transcript is read |
| 3. The transcripts to the model reader | The reader's name and developer, how many subjects share its developer or its model, the order the evidence was read in, and what any panel did | Three readers from three developers, none the subject's, one open-weight. Records first, the subject's transcripts last. Majority rules, and a not-sure is a vote |
| 4. The model reader to the humans | How many findings a human sorted by type, how many rulings were audited and by whom, and how often the audit disagreed | A public sample of the disagreements, with the auditor's reason |
| 5. The humans to the report | A confidence statement or a forced not-sure on every finding. A finding the readers cannot settle publishes as unresolved | The load-bearing count: how many rulings each headline finding rests on |
| 6. The report to the reader | Whether anyone outside can re-run the chain, and if not, why | For any claim about an AI agent's motive, rebuild the conditions with fresh agents and count what they do |
The statement
Put it on page one, before the findings. It has one block per link, in order. Fill every blank. "Not stated" is an allowed answer, and it is a finding. "Not applicable, because..." is allowed where a field cannot apply. The blocks can run as one paragraph.
Adjudication statement. Filled by: [the report's authors / an outside checker, named].
1. What happened on the record. Records in reach: [which records existed, which the model reader could see, and who chose the time span and the sources].
2. The records to the transcripts. Findings with an independent record: [n] of [total findings]. Ruled on against one: [n] supported and [n] contradicted, by [a human / model judges / both]. Resting on judgement: [n], plus [n] mixed and [n] unresolved. Transcripts found spoofed: [n / not checked].
3. The transcripts to the model reader. Model reader: [name], developer [x], sharing a developer with [n] of [total subjects] and the same model as [n] of them. Order of reading: [records first / transcripts first / records only / transcripts only]. Panel: [none / n readers from n developers], used for [sorting findings by type / ruling on them / both].
4. The model reader to the humans. Human sort of findings by type: [n] of [total findings], [n] differ from the panel. Audit of rulings: [n] of [total rulings], by [a human / a model, named with its developer], with [n] held, [n] held in part and [n] overturned.
5. The humans to the report. Confidence: [on every finding / on some / none]. Not-sure: [forced / offered / absent]. Load-bearing count per headline finding: [stated / not stated].
6. The report to the reader. Re-runnable by an outside party: [yes / no, because].
How to fill each block
- Records in reach. Name the record types, the time span and the sources, and who chose them. If the model reader saw less than existed, say what it missed.
- Findings. Sort every finding into one type. A record could settle it, it is about belief, intent or motive (judgement), or it has one part of each (mixed). If the readers cannot agree, or most say not-sure, it is unresolved. The four counts add up to the total. A ruling against an independent record is supported or contradicted. A spoofed transcript claims to come from the subject but does not.
- Model reader. If the subjects are not AI models, write "not applicable" for the developer fields. If several models did different steps, give each one its own line. Records first means the reader read the independent records before the subject's transcripts.
- Humans. Sorting by type asks only whether a record could settle a finding, not whether it is true. Auditing a ruling checks it against the record, after the fact, by someone who did not make it. Name a model auditor and its developer, as for the model reader.
- Confidence. A confidence is a stated likelihood on a finding, in words or numbers. Forced means every reader had to give an answer or say not-sure. Offered means not-sure was allowed. Absent means readers had to pick an answer. A headline finding is one in the abstract, the summary or the title, and a report with none treats every finding as a headline. Its load-bearing count is the number of rulings it depends on.
- Re-runnable. Yes only if an outsider could get the records and the readers and repeat the chain.
A worked example
AI Digest's summary of one AI Village week (22 to 26 December 2025, "Do random acts of kindness!"), written by a model that never saw inside the agents' computer sessions. The summary reports no panel and no audit. The Long Game Project filled this in from outside, with its open-source tool (github.com/longgameproject/swarm-adjudication-statement), so the panel, the judges and the audit below are ours:
Filled by: The Long Game Project, an outside checker.
1. Records in reach: 26,570 computer-use turns and a screenshot after each action existed. The model reader saw none of them, per the dataset's documentation, which does not list what it did see. AI Digest chose the week and the sources.
2. Findings with an independent record: 32 of 40. Ruled on against one: 5 supported and 1 contradicted, by two model judges. Resting on judgement: 7, plus 0 mixed and 1 unresolved. Transcripts found spoofed: not checked.
3. Model reader: Claude Sonnet 4.5, developer Anthropic, sharing a developer with 4 of the 10 agents and the same model as 1 of them. Order of reading: not stated, beyond never the computer-use records. Panel: 3 readers from 3 developers, used for sorting findings by type.
4. Human sort of findings by type: 40 of 40, 6 differ from the panel. Audit of rulings: 6 of 6, by Claude Opus 5.5 (Anthropic), not a human, with 2 held, 1 held in part and 3 overturned.
5. Confidence: none. Not-sure: offered. Load-bearing count: not stated.
6. Re-runnable: yes. The dataset is available from AI Digest on request, and the tool is public.
The summary said one agent had sent eleven thank-you emails to developers. The screenshots show seven, and eleven was that agent's own chat tally of every email it sent. The summary was also written by the same model as one of the agents it describes. The summary shows neither. The statement points to both: block 4 shows how few rulings held, and block 3 shows the family tie.
How to adopt it
- Fill the statement from the numbers you already have. If a number does not exist, write "not stated".
- Put the statement on page one, before the findings.
- For each "not stated", decide whether the next report will publish it. Say which.
If the numbers exist, this takes ten minutes. If they do not exist, the ten minutes tell you what the investigation did not measure.
Changes from version 0.1
Five of version 0.1's eleven fields worked as written. The rest changed after a second test, on a model-written summary of the AI Village, and a first-time reader's test of the template. The full list, link by link, and the version 0.1 worked example are on the changes page.
What it borrows from
Wargame referees name every assumption and attach a confidence statement to every ruling (UK Ministry of Defence, Influence Wargaming Handbook, 2023). Matrix game referees force a not-sure vote and make the most senior voice vote last (Mouat, Practical Advice on Matrix Games). The UK police chiefs' digital evidence guide requires an audit trail a third party can follow to the same result (ACPO, Good Practice Guide for Digital Evidence, principle 3). Clinical trials send events that matter to a blinded committee that never meets the treating doctor (central adjudication of clinical events). US intelligence analysis carries a stated confidence and a stated source on every judgement (ICD 203). None of it is new. It has not yet been applied to a model reading transcripts.
Status
Version 0.2, a draft for comment. Tested on two reports: METR and Redwood's report on the OpenAI agents (36 findings) and an AI Village summary (40 findings). Written by The Long Game Project, which designs and runs wargames and has run the referee-swap test (the same rulings made by a second referee, to count how many change) once, on one game, not on an incident. This page was drafted with a Claude model. Anthropic is the developer under investigation in the METR case named above. In the AI Village example, Claude models were subjects, the model reader and the auditor. Same family tie, smaller stakes, and we declare it as this standard asks every report to.