500K Preflight

The document begins by stating that this runtime did not independently recompute the official 500K output. Before scaling up, five steps must be completed in order: lock down environmental variables like Sage/LMFDB/Git SHA/descent backend; perform an exact replay on conductor<150; perform a first-failure replay on the 13 curves present only in the old version; perform an exact rejection replay on the four discrepancy curves from document 07 (upcoming); and only then run conductor<500000. After the 500K run is completed, seven machine-readable artifacts must be output: passed.csv, failed.csv, unknown.csv, predicate_trace.jsonl, descent_certificates/, run_manifest.json, and hashes.txt—the document specifically emphasizes that unknown.csv must not be discarded, otherwise a timeout or backend failure would be misread as a mathematical rejection, which is exactly the concrete implementation requirement of the S3 clause in document 10. Finally, it defines the precise standard for "success," explicitly not a vague judgment like "the final count is very close," but four things holding simultaneously: final set exact (not approximate), stage counts reproducible (counts at each filter stage can be reproduced), adversarial corpus stable (results for the adversarial corpus from document 09 are stable), and certificate semantics stable (certificate semantics do not drift across versions).