Vérification

DEMONSTRATED · shipped executable checks

verify_sample.py reads all three JSONL splits, applies SCHEMA.json, checks unique IDs and synthetic provenance assertions, and runs independent label validators. GameNav uses Bellman-style relaxation and path legality/cost checks. GameCoord uses joint-state A* and synchronous plan checks. These implementations are included in each archive.

Negative tests and limits

For each of 100 records per sample, the verifier alters one cost or makespan and requires rejection: 200/200 tested mutations across both samples. This is targeted negative coverage, not exhaustive adversarial robustness or human certification. Null reachability labels are checked by the independent oracle.

Integrity and provenance

SHA-256 is calculated on complete archive bytes. CHECKSUMS.sha256 covers package files except itself; MANIFEST.json excludes itself and CHECKSUMS.sha256 to avoid circular hashes. The provenance says original_procedural_synthetic, deterministic_computational_rules, no external records, no personal data and no LLM ground truth. Read PROVENANCE.md and SOURCE_LICENSE.json inside each archive.

EXPERIMENTAL / PLANNED

Broader client environments, richer failure/recovery corpora and continuous evaluation require their own execution and rights evidence. No result on these 200 records certifies another environment. These docs do not assert public support for those planned extensions.

Release evidence