Post-hoc alternate blind reviews after invalid output

Registered before these calls. Six completed attempts in batches 31 and 32 have a valid DeepSeek rating and a Qwen response that completed but failed the format check. The manifest, SHA-256 65286818dfddd2188d88369d2dcf68711a72ba453610603fde270f17ce2add26, freezes the selected attempts, source and anonymized packet hashes, randomized order and model limits. Missing or failed/timeout reviews are excluded; original ratings and failed outputs remain separate.

A new gpt-6.1-sol blind review is planned for each of the six attempts. If that review and the original DeepSeek rating are valid but disagree, gpt-6-astra is asked to adjudicate. Each stage stops on provider failure or missing usage; only an explicit HTTP 503 can receive bounded retries, never a timeout. Requested output is 1,200 tokens per call, not a hard gateway spending limit. A candidate with a missing or invalid new response remains missing or invalid.

This selection and model substitution are post hoc. The adjudicator has the solver’s model label and is not independent of it; all model labels share a proxy with unverified provenance. None is a human reviewer, and format validity is not proof of factual correctness or project acceptance.