Benchmark Codex OFF vs MARGINAL with fail-closed evidence

[SignalLayerLabs]

Outcome

Adds the audited Codex 0.147 OFF/ON adapter, frozen SWE-bench Lite smoke protocol, six measured inference trajectories, official verifier evidence, and an evidence-first README/site surface.

Measured smoke

• 3 matched tasks, one repetition per task
• resolved: 0/3 OFF and 0/3 ON
• effective tokens: 1,098,747 to 824,839 (24.93% fewer observed)
• effective latency: 593,106 ms to 565,768 ms (4.61% lower)
• governance: 0 external tokens, $0, 7,058 ms
• evaluator status: pass_through; not eligible for a support claim
• no applied denies in the three real trajectories

Verification

• authoritative SWE-bench 4.1 Docker verifier: 3/3 completed per lane, zero infrastructure errors
• Modal audit run 31474500980: symmetric image-build error retained and excluded from scoring
• infrastructure error_ids and incomplete_ids now fail closed
• zero verified successes can no longer produce a supported result
• local gate: 337 tests, Ruff, mypy, protocol validation and JavaScript syntax all green

Scope

This is an exploratory integration smoke, not a general token-saving performance claim. The 20-task repeated canary remains behind the preregistered promotion gate.