← Back
📝 Notes / reports / phase3a-pubchem-identity-discrepancy-20260919

📑 Contents

PubChem identity-count discrepancy audit — 2026-09-19FindingDecisionCaveat

PubChem identity-count discrepancy audit — 2026-09-19


Finding


The source expectation records 231,755 declared unique identity14 values, while the built index contains 231,749 computed unique identity14 values: a net difference of six.


The discrepancy is not six missing valid structures. The builder recomputes the authoritative scorer identity from SMILES and records 614 source-row identity mismatches. The direct set comparison from the source audit found:


  • declared source identities: 231,755
  • computed/index identities: 231,749
  • declared identities absent from computed/index set: 578
  • computed/index identities not present in declared source set: 572
  • net: 578 − 572 = 6

    • The index therefore intentionally follows the pinned local identity algorithm rather than trusting the source-declared identity field. This is consistent with the builder's identity_mismatch: 614 diagnostic and the verifier's integrity result.


    Decision


    Retain the current index. Do not treat the net six-count difference as six catalogue-coverage misses. Preserve both declared and computed identity counts in provenance, and report the 578/572 set differences if this artifact is used in a later audit.


    Caveat


    The expensive full recomputation was terminated after the set-difference result had already been established from the source/index audit; no source rows were modified and no production artifact was changed.