Permit validators still over-flag after adding a vague be-conservative instruction. Why?

Permit validators that still over-flag after a vague "be conservative" instruction illustrate that conservatism slogans are not categorical controls. Vague be-conservative wording fails compared with explicit report-and-skip criteria because "conservative" does not say which finding types to emit, which to suppress, or what evidence is required. Explicit categories with examples constrain false positives; fuzzy caution language often increases them as the model hedges by flagging more.

The claim that vague conservatism always outperforms explicit categorical criteria fails both as prompt-engineering practice and against the stem's residual over-flagging. Saying over-flagging can only be fixed by deleting the validator fails because the remedy is better criteria, not abandoning validation of permits. Asserting that conservatism instructions automatically encode categorical examples fails: adjectives do not smuggle in report/skip lists or worked examples unless you write them.

Exam caveat: replace conservatism fluff with categories; do not confuse "fewer findings" as a goal with "skip security"—skip low-value noise, keep high-severity permit defects. Operational check: remove be-conservative boilerplate, add report/skip criteria and examples, re-run the same permit corpus, and compare precision/recall rather than raw flag count alone.