Menu
← Back to the standard

The adjudication statement: changes from version 0.1

Version 0.2, October 2026. The Long Game Project. Back to the standard.

We tested version 0.1 on one report. Version 0.2 comes from running it on a second kind of report, a model-written summary of the AI Village, and from a first-time reader's test of the template (October 2026). Five of version 0.1's eleven fields worked as written. By link:

  1. What happened on the record. Version 0.1 had no field here. Added "Records in reach", and a "Filled by" line so an outside checker says so.
  2. The records to the transcripts. "Verified" counted two judges agreeing as verification. It now gives supported and contradicted separately and says who ruled. The template now carries the spoofed count from the table, and "unresolved".
  3. The transcripts to the model reader. The developer field counts shared developers and shared models across many subjects. "Records only" and "transcripts only" join the order of reading. The panel says what it did. One pair of words, records and transcripts, replaces logs and traces.
  4. The model reader to the humans. One field mixed two checks: sorting findings by type, and auditing whether a ruling holds. They are now two fields, and a model auditor is named with its developer.
  5. The humans to the report. Confidence gains "on some". "Headline finding" is defined. "Unresolved" no longer assumes a three-reader panel.
  6. The report to the reader. The field is unchanged. The better version now applies to claims about an AI agent's motive.

Version 0.1's worked example, METR and Redwood's report on the OpenAI agents (26 August 2026), filled from the authors' own disclosures and our count of its 36 core takeaways:

Findings with an independent record: 26 of 36. Findings verified against one: not stated. Findings resting on model judgement: 7, plus 3 mixed. Model reader: GPT-5.6 Sol, developer OpenAI, same as the subjects' developer, same model as about 5 per cent of them. Order of reading: traces first. Panel: none. Human audit: attempted, fraction not recorded. Confidence bands: none. Not-sure: absent. Load-bearing count: not stated. Re-runnable: no, the corpus stays at OpenAI.