A larger campaign still could not identify support.

Live frontier-provider science

The rebuild was live—not merely an offline qualification.

The accepted campaign used Anthropic Claude Fable 5 and Google Gemini 3.6 Flash, with 888 planned and resolved coordinates: 444 per provider, 864 MAIN observations, 24 mechanical-control observations, and 288 planned matched triples. It completed with zero technical retries, missing coordinates, duplicates, or unexpected coordinates.

The scientific result is bounded and exact: both provider arms returned INSUFFICIENT_EVIDENCE, with no authorized provider-specific support claim. Only 66 matched units were fully valid; 222 required the frozen pessimistic coding rule. The mechanical XC construct gate passed for both providers, but that did not override the MAIN coverage problem.

Observed coverage burden

Scientific result versus later diagnosis

The accepted confirmatory result ends at INSUFFICIENT_EVIDENCE. A separately labeled post-confirmatory analysis found that all 66 fully valid units showed nuisance movement; Google had 29 target movements, none selective, while Anthropic had no target movements. Those observations help diagnose the design, but they are not a replacement confirmatory result.

What it changed

Experiment 2 showed that more calls and stronger custody do not automatically repair identification when refusal, truncation, and nuisance movement dominate matched observations. Experiment 3 therefore moved from a target-minus-control scalar toward ordered dose response, refusal-aware common denominators, and joint partial identification.

Machine-readable public summary · ← Experiment 1 · Continue to Experiment 3 →