← Back
📝 Notes / reports / g1-acceptance-20260919 / G1-ACCEPTANCE-REPORT

📑 Contents

G1 Local Acceptance Report — 2026-09-19ExecutedFrozen evidence hashesResult interpretationGate decisionNext authorized development track

G1 Local Acceptance Report — 2026-09-19


Executed


  • Maintained contract/similarity tests: 16 passed.
  • Rebuilt-index provenance audit: passed.
  • Regime-A policy/mask audit: passed.
  • All five Regime-B mask audits: passed.
  • Fresh runtime submission validation: passed, 400/400 rows, zero errors.
  • Independent candidate identity audit: passed; every row has at most 25 raw candidates, no invalid SMILES, no duplicate scorer identities.
  • Offline inference smoke test: passed on an isolated artifact view with renamed and reversed molecule IDs (smoke_C, smoke_B, smoke_A); 3/3 rows produced in the requested order.

    • Frozen evidence hashes


    See the machine-readable hash bundle printed during the acceptance run. Key inputs:


  • data/train.parquet: 9423f90ec08b941a9646b535bf0606183e86ff0c5a297978f5a70693d41bcd98
  • data/test.parquet: 2fa29c4aa06e93b48e0cd6619fd2de3f4e956f0019d242141678b7287f2b43e0
  • reports/phase1-g1-all-20260918.json: 34dd4a40db9178a3900a5e319c1d08f83da01bf9ff1a9a6157fe5d62bfc221d4
  • submissions/phase1-retrieval-rebuilt-20260918.csv: 83e3209f445f76e91ba59117fb3b6dd656699a1f1d45a2370d04432bcd1ae6d4

    • Result interpretation


    The all-spectrum report is qualified as strong local retrieval evidence, not an official competition score:


  • Library ceiling: Top-1 83.53%, Top-5 89.57%, Recall@25 90.16%, pool coverage 90.37%.
  • Regime A proxy: Top-1 88.0%, Top-5 98.8%, Recall@25 100%, pool coverage 100%.
  • Regime B: target coverage 0 by identity-wide leakage-control design; not a quality score.

    • Gate decision


    Qualified local G1: PASS WITH CAVEATS.


    The local retrieval baseline and packaging contract are reproducible and accepted for subsequent offline development. The following remain external/unresolved and must not be presented as closed:


    1. Official CASMI scorer executable/reference-vector parity is unavailable.

    2. Regime A certifies only the local exact-available-field proxy, not acquisition lineage.

    3. Regime C true novelty evaluation is not available; no C score is claimed.

    4. The local library-ceiling result is optimistic diagnostic evidence, not hidden-test generalization.


    No Kaggle/public submission or model training was performed by this acceptance run.


    Next authorized development track


    Proceed to Phase 3A/3B structure-only and formula/mass-gated candidate-pool work, with pinned source provenance and leakage-safe evaluation. Do not start large learned-model training until that candidate pool and non-neural baseline are frozen.