Observed arrival · 2026-10-16
Impeach: a court for benchmark scores
Impeach presents a public process for challenging model benchmark scores with rewritten questions, verbatim-recall checks, and inspectable records.
- For
- Model publishers and evaluators checking benchmark claims
- Worth noticing
- The displayed Bradbury rehearsal is labeled scripted and local, not a chain record; raw counts and recomputation files are published.
Field notes
A claim begins with a score, a bond, and hashes of its test items; challengers stake against it, and rewrites must pass two referee framings before admission. The page exposes a sample answer sheet, certificate and case-file downloads, plus a registry that marks revealed items spent to discourage reuse. Its displayed Bradbury/Llama rehearsal is explicitly described as local and scripted rather than a chain record; the site also provides source and contract specifications for rebuilding the published verdicts.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue