6,814 blind judgments across 33 questions
Research question
Which test cases are unambiguous, and where do reviewers disagree?
Approach
Models review a larger synthetic test set without seeing the intended answers. Disputed cases are examined separately.
What is available so far
A documented GPT-6 run contains 6,814 judgments, including 1,359 marked uncertain. A later list records 489 disputed cases from the review collection.
Limits of the findings
These counts describe different analysis stages. Model majorities do not replace a human reference; a complete, cleaned overall evaluation is not available here.
Next step
Resolve disputed cases and identify a fixed, separate evaluation set.
Basis of this account
- Blind-review report dated 4 October 2026
- Disputed-case list dated 5 October 2026
This page summarises the documented state of the work. A full standalone presentation of the results has not yet been prepared.
The data snapshot refers to the underlying investigation, not the date of this overview.