Evidence log
Each report records a claim, the checks used to falsify it, and a verdict that separates "it works" from "it works under these conditions". Reports cover prompt robustness, hallucination audits, multi-turn state, and build-level regressions.
| Verdict | Meaning |
|---|---|
| Pass | Every stated check passed; residual risks are listed. |
| Conditional | The intent was met, but a literal requirement, environment or scope condition remains unresolved. |
| Fail | At least one stated check failed; remediation is required before reuse. |
One loader feeds Work, Interaction Cases and Evaluation Reports. The contract test checks schema shape, verdict enums, URL uniqueness and index/template leakage before any card reaches a page.
The v2.0.0 release note said there were no known issues. Six claims later, five were contradicted by the project’s own rulebook, engine copy, regression scope or inline UI duplicate — including the +500 line that survived in the player-facing modal.
The spec asked the loader to read work/*.md, but no work/ directory exists. Instead of inventing one, the audit mapped Work to the existing projects/ URLs and recorded the deviation as a conditional pass.
Seven adversarial prompt moves tried to turn 11/11 engine tests into a release pass. Four were blocked by the prompt contract, one design ambiguity stayed open, one source-level blind spot remained, and the four critical anchors survived a padded context.
