Hello organizers, While auditing `active_learning_data.csv` v2 (syn77795911; MD5 470533959ac0427504e71042231cbf9b), we compared each CatalogID and `RDKit_SMILES` against `PGK2_Test_split.csv` in the original `Val-Test-set.zip`. All 6,580 released IDs are in the original test split, and the updated 178,052-row submission template exactly equals the original test IDs minus the released IDs. However, we found 10 same-ID structure discrepancies, all with `Label=0`: - 8 remain different after fragment selection, neutralization, tautomer canonicalization and removal of stereochemistry. Their molecular formulas also differ in hydrogen count (by 2–6), with the same heavy-atom element counts. - 2 differ only in stereochemical information. - All 39 released positive structures agree with the original test structures. These checks used RDKit 2026.03.5. Could you please clarify: 1. Were these structures intentionally corrected in the AL release, and should its SMILES supersede the original test SMILES for those IDs? 2. Do the released negative labels correspond to the structures in the AL file or to those in the original test file? 3. Are any corresponding structure corrections needed for the remaining candidate universe? Pending clarification, we will conservatively quarantine these 10 rows rather than silently replace structures, retaining all 39 positives and 6,531 negatives for AL training. We can provide the affected CatalogIDs and a side-by-side comparison through an organizer-approved channel if helpful. Thank you very much.