The temptation to satisfy the proxy
When a detector flags valid prose, the fastest green is a small rewrite. The check passes and the dashboard quiets, but source has changed to match an instrument bug rather than meaning.
This is especially dangerous for memory. Removing placeholders, names, or intentional repetition that the lens cannot model deprives the next soul of information. The test passes while the product purpose fails.
Verify the premise first
Read the candidate and decide whether the claimed class is actually present. On a false positive, restore content and narrow the predicate. Prefer an exact description of the legitimate shape over a broad ignore.
If an English X triggers a Korean shorthand detector, allowing every X hides real defects. Classify placeholder position, code block, or mathematical symbol and exempt only those contexts.
Keep the incident shape as a fixture
After repair, preserve a reduced version of the real false-positive fragment. The valid sample must pass while a neighboring invalid sample still fails. With only one side, disabling the detector also appears green.
Place samples on both sides of numeric thresholds and retain raw syntax for parser failures. Generalize the lesson without erasing the unusual shape that exposed it.
State instrument limits in the UI
No detector is perfect. Candidate cards should show why they were flagged and known limitations so users read the result as evidence rather than an oracle.
When reporting false-positive rate, attach period and sample. A clean run does not prove perfect precision; show how many reviewed candidates were confirmed.