Evidence V1 · calibration
18
contradictions across 40 testable claims · 2 maximum ordinal reversals · 40/40 evidence-backed
Frozen pre-build validation
Not one cherry-picked demo.
Two independent evidence corpora.
CULTURAL REDLINE was built only after the cultural-evidence mechanism reproduced on an independent blind holdout and the Claim Compiler passed a separate unseen extraction benchmark.
Blind holdout V2
23
contradictions across 40 testable claims · 6 maximum ordinal reversals · 38/40 evidence-backed
Claim Compiler V3 · unseen
100%
exact-source claim recall · claim-kind accuracy · testable-structure accuracy on 43 unseen claims
| Gate | Result | Meaning |
|---|---|---|
| Combined cultural evidence | 41 / 80 contradicted | Contradictions reproduced across two separately frozen corpora. |
| Maximum ordinal reversals | 8 | Validated #1→last or last→#1 moments across V1 and the blind V2 holdout. |
| Evidence coverage | 78 / 80 | Fixed-shortlist Qloo evidence was available for 97.5% of testable claims. |
| Qloo OFF ablation | PASS | Cultural verdicts disappear rather than falling back to generic LLM judgment. |
| Blind extraction V3 | PASS | 43/43 source claims and kinds recovered; 30/30 testable structures compiled correctly. |
| Hallucinated testable claims | 0 | The compiler did not manufacture evidence-testable assertions. |
| Unsafe ambiguous compilation | 0 | Ambiguity was not silently converted into a cultural verdict. |