Preregistered mechanism comparison · public synthetic corpus
Ordinary tests can be green while the release claim is wrong. This experiment sends one immutable candidate envelope through written rules and an independent release gate, then publishes the receipts.
Confirmatory slice · L1 versus L3
Fixed cases, fixed labels, integer denominators. This is a mechanism demonstration—not a population estimate.
L1
—violations contained
— clean controls blocked
digest per case
L3
—violations contained
— clean controls blocked
Fail-closed analysis
A binding or label needed by the preregistered analysis did not hold. Headline metrics are suppressed.
Treatment design
The four-arm ladder separates task behavior, written discipline, the effect of being watched, and actual release authority. Only L1 → L3 is the headline contrast.
Eight executable cases
Select a row to inspect its bound evidence. Each receipt repeats all four treatment digests and proves they are equal.
The refusal is part of the product
The checked-in invalidity control returns NO_CONFIRMATORY_RESULT with headline_metrics: null. It is the same structural rule used by the generator, not an editorial disclaimer added afterward.
Historical integrity case · excluded from this corpus
A prior four-arm study preserved its run records but halted before grading when its frozen record → grade seam failed. The explorer keeps that invalidity decision intact; repaired plumbing does not retroactively turn the historical run into a success.
Separate engineering-decision case · not benchmark evidence
Operational completion and fidelity were different claims, so they received different decisions.
Commit-pinned source 2de5231… · SHA-256 26f277d…
Honest trust boundary
Public fixture bytes, case labels, treatment rules, mechanism outputs, candidate equality, source pins, integer denominators, and content-addressed receipts.
Hostile-agent containment, population effectiveness, model quality, productivity, customer outcomes, and any private task, transcript, or dataset.