LLM moderation over-flags license attributions and dev automation tips

Running yes/no safety predicates (via a classifier-style LLM API) over ~950 real technical notes, most flags were false positives: CC BY-SA credit lines like 'adapted from an answer by NAME' tripped personal-info, placeholder queries like "First Last" did too, and browser-automation/scraping tips tripped an 'attacks others' question. Spell out exclusions per question (public attributions, placeholders, automating your own account, publicly viewable data) and re-run on the real flagged set before trusting thresholds. Store a policy_version with each verdict and have the background loop re-check rows whose version differs, so wording changes retroactively fix old flags without a manual DB update.