Check the target before you propose a fix

A signature that matches passing rows is a lead, not a defect. This skill exists because one project's review named a naming-output pattern (a prefix emitted without its locant) as the top fix target, and measurement then showed:

• 30 hits, 29 failures — the 30th row already round-tripped correctly in the shipped config, so the flagged construction is legal in one sub-family
• three unrelated causes, not one: missing-locant on an acyl chain (~15), a glycosidic-oxygen case (~13), a raw input string emitted verbatim as output (2)
• the glycosidic sub-case is a double-counted feature — the parent structure was named as if a substituted oxygen were a free hydroxyl, so a locant-only fix would leave the output wrong

The lesson generalizes past naming: a whole iteration was scoped against "one defect" that was three, one of which was not a defect at all.

The procedure

1. Resolve the target to an explicit row list from a committed eval/benchmark run. Not a description, not a cluster name — the actual inputs (SMILES, molecule IDs, or whatever your unit of analysis is).

2. Recompute the signature over ALL emitted rows, not just failures. This is the whole point. A signature computed inside a failures loop is structurally incapable of seeing its own counterexamples — a clustering script that only iterates failing rows has exactly this bug.

3. Report four numbers:

4. Decompose by cause. Inspect the actual outputs (names, labels, values). If the matched rows split into groups with different mechanisms, the target is N targets: re-scope to the largest one and return to Step 1.

5. Check the outcome spread. A single defect class does not usually span parse_fail + wrong_structure + exact_match. If the hits span 3 or more outcome types, return to Step 4: there is more than one cause.

Refusal conditions

Report NOT A SINGLE DEFECT and stop if:

• precision < 1.0 — the signature matches rows that already succeed. Call it a lead and name what distinguishes the failures from the passes.
• n < 10 — too small to be worth a task, and too small to see a second cause.
• the hits decompose into ≥2 mechanisms — re-scope to the largest single mechanism and re-run this check on it.
• the hits span ≥3 outcome types — almost certainly multiple causes.
• the signature is a string pattern on the emitted output and no structural cause has been identified. Output patterns are symptoms; the same wrong output arises from unrelated causes.

Size from the named inputs, not from a tag

Size a defect class only from the actual inputs, each traced to its cause; a count of rows that match a tag or substring is not a size. Projects repeatedly over-count 15–50× by counting tags: a curated "~100 candidate defects" list that turns out to have 2 genuine instances on re-derivation; a static sweep's "12 live defects" that turns out to be 4 once each case is traced.

Related

• spy-site — once the target survives this check, prove the code site is on the path.
• cluster-failures — produces the candidate targets this skill validates.
• Your project's fix-methodology / engineering-log doc, if it has an equivalent invariant.