A PASS was retracted as contamination four minutes after it was reported

What we expected

Extending the pocket-knockout panel with 43 additional compounds (mutant arm only) should give a cleaner read on whether Boltz-2's signal survives pocket destruction.

What happened

The merged panel reported a PASS — after-arm AUC 0.5156 [0.3775, 0.6495], upper bound finally under the 0.70 memorisation bar. Checking the class labels afterward found the panel was not what it claimed:

• Compound 78c (CD38 reference) — a different target entirely
• theobromine, apigenin, 2 glycosides — natural products, activity unknown
• 2 UUIDs, 2 bare numeric ids — unidentifiable
• 12 stereoisomers of one alkaloid — ONE molecule counted twelve times

All were labelled “active” by the rule 0 if name.startswith("decoy") else 1 — a rule that a morning audit the same day had already flagged as fragile, and that had not yet been fixed. Verified-only re-analysis: n=68, auc_after 0.5761 [0.4249, 0.7232] → INCONCLUSIVE.

Why it happened

The twelve stereoisomers — one molecule counted twelve times, more than half the added class — dropped hard under the knockout (0.85 → 0.20–0.46), pulling the after-arm AUC toward chance and manufacturing the collapse.

What changed

“A contaminated panel did not merely add noise — it moved the verdict, in the direction most flattering to this project's headline number.” Labels now require measured evidence: a measured potency against the specific target, a curated holdout, or membership in a named KNOWN_ACTIVES allowlist. Everything else is excluded with a printed reason. Commit d2050b1.

“Seventeenth defect of the day, and the first one that reached a stated conclusion before being caught. Every earlier one was found before it produced a number I published. This one I announced.”

Lab notebook, line 6383.