src/lib/ai/quality-gates.ts:
- Read bb.banned_terms with the ban_level column so admins can mark
a term as hard (instant fail) or soft (score nudge).
- Hard hits join the hard-coded AI_TELLS list to instant-fail the
rewrite (status=failed_quality, regenerate).
- Soft hits each contribute +0.05 to the heuristic AI-detector score,
capped at +0.30, so a borderline-slop article can be tipped over
the 0.65 threshold without single-handedly being rejected.
- Multi-word phrases are matched as substrings; single words use a
boundary regex that handles hyphenated terms cleanly.
- QualityGateResult now exposes softWordsFound alongside the existing
bannedWordsFound; failure_reason text includes the soft list when
the score trip is what failed.
44 terms seeded in bb.banned_terms (23 hard hype words like
"revolutionary", "industry-leading", "world-class", plus 7 PR-speak
openers like "pleased to announce"; 21 soft like "innovative",
"leverages", "robust", "transform").
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>