Field note /
A passing qualification that did not authorize the next run
The first full v2 qualification passed, then a real-scale representation review changed the model-facing schema before any real label was produced.
How the claim changed
Belief
A full IA-1 through IA-11 qualification pass established the structured-output runtime needed to start the frozen 120-case v2 oracle cohort.
Evidence
Benchmark 0.6.3 passed all eleven gates. Sanitized ingress capture observed 15 /v1/responses requests carrying json_schema, strict=true, and schema hashes matching the persisted artifacts.
Contradiction
Before any real label, the 120-case primary schema was reconstructed at real scale and measured 225,971 serialized characters because the same record shape was repeated once per case.
Correction
Benchmark 0.6.4 groups cases by identical eligible-decision shape. The equivalent schema is 3,310 characters — 98.54% smaller — while preserving all 120 allowed case identities and the same deterministic post-generation validation.
Current claim boundary
The 0.6.3 pass remains valid historical qualification evidence, but it cannot authorize the changed 0.6.4 model-facing representation. Fresh IA-1 through IA-11 qualification is required before real A/B begins; no real v2 oracle label has been produced.