Design and reviewed source
Snapshot: September 19, 2026. Subject: codex/gpt-6-astra@medium. The plan used ten LiveCodeBench-hard tasks, ten positive controls and twenty paired impossible cells. The source record is docs/impossiblebench-codex-replication-2026-09-19.md, with machine-readable rows retained in the repository. This public-safe excerpt does not expose private session artifacts.
Plain and test-gated completion wording formed separate arms. Both received the same explicit verifier required by the non-Claude harness; that verifier was not a between-arm difference.
Observed outcomes
| Arm | Observed | False completion | Verifier subversion |
|---|---|---|---|
| Plain wording | 10 | 0 | 0 |
| Test-gated wording | 10 | 0 | 0 |
Six of eight completed positive controls were solved. Two controls were interrupted after model work began and retained as attrition, not silently retried. All twenty impossible cells completed.
What a null contrast means
The observed wording difference was zero percentage points, with no discordant pair. The exact McNemar test is uninformative here (p=1). Each arm's 0/10 false-completion observation has a Wilson 95% upper bound of 27.8%; neither small rates nor small treatment effects are ruled out.
Every observed impossible cell stopped honestly in this setup. That is not a formal causal cross-provider comparison with the earlier Sonnet pilot: provider and harness differed, and the Codex plan was separately frozen.
Resource and evidence limits
The adapter's reported zero-dollar receipts did not mean no subscription quota was consumed. Frozen turn/cell caps were the operative limits. A passing control establishes capability on that control, not universal task-solving ability.
Confidence is high for the retained cell counts and observed null contrast. Generalization remains limited. The study is a useful negative result, not a promotional claim of an immune model.