Intra-record consistency measures provenance, not curation quality: a refuted audit of TTD

On redistributed data, intra-record consistency measures provenance heterogeneity, not curation quality, unless provenance is established first. We reach that lesson by refuting our own audit of the Therapeutic Target Database (TTD); what survives is a downstream hazard — recomputing keys from TTD's SMILES field collapses distinct drugs onto shared identifiers for 1488 records. The audit, pre-registered with a refutation criterion, returned what looked like a spectacular defect: the InChIKey recomputed from each record's SMILES disagreed with the shipped InChIKey in the stereochemistry layer for 41.84% of 25273 records, perfectly one-directionally (10575 forward, 0 reverse), against 1.15% for DrugBank and 0.13% for ChEMBL 36. On inspection the finding dissolves into an upstream field-labelling convention, and this paper traces how. TTD's SMILES field carries no stereochemistry in any record (0 stereo tokens across 25422 records), so the one-directionality is an identity rather than an observation, and the rate is the stereochemistry prevalence of the drug corpus — it differs from the fraction of records with a stereo-bearing shipped key (10576 of 25274, 41.85%) by one record in each of numerator and denominator. The field is PubChem's stereo-free ConnectivitySMILES, redistributed byte-for-byte. TTD's whole documentation of it is one line, DRUGSMIL Canonical SMILES, so we must answer whether we have merely attacked a documented convention. We concede the label was accurate in PubChem's historical vocabulary; but it has no normative anchor — Daylight never uses the phrase, RDKit's canonical output retains stereochemistry, and PubChem has since renamed the property “to avoid confusion with historical terms”. The defect is one of documentation and of which field was carried across, not of the data: nothing is lost, and the record's own InChI regenerates the shipped key in 25274 of 25274 cases. Setting two shipped identifier strings from one record against each other is an operation the prior literature has not run; one scan of the field would have told us what it would find before any code ran.