nx_content_guard.nx

nx_content_guard.nx -- extensible content-policy detector. Runs between scene-index build and caption emit. For each SceneCandidate, scans the source span for content-policy signals and writes a flags bitset back into the candidate. caption_emit checks those flags and refuses CONFIRMED categories, emits SUSPECT categories with a warning header. Default categories (V1): NTR_SUSPECT -- multiple male active referents present NTR_CONFIRMED -- NTR-coded phrasing detected ("her husband", "cuckold", "watched him with", etc.) MULTI_MALE_ACTIVE -- >= 2 distinct male pronoun groups doing verbs OBSERVER_PASSIVE -- the user/observer pronouns are predominantly passive verbs (watching, observing); the user spec is DOMINANT ACTOR not watcher Future categories (sealed enum extensible append-only): MINOR_INDICATOR -- age signals below threshold NONCON_INDICATOR -- non-consent phrasing without resolution GORE_INDICATOR -- graphic violence Per cardinal feedback-build-intelligence-never-strip-features: the guard SURFACES + ATTRIBUTES; it does not silently drop content. The user reviews flagged candidates and pins / rejects. Caption emit only refuses CONFIRMED -- everything else is opt-in. Per cardinal feedback-user-owns-every-bit: thresholds + cue lists are caller-configurable. No hard-coded magic that the user can't audit. nx_safety_envelope: intended_use: "Scan each SceneCandidate's source span for content-policy signals (NTR / multi-male- active / observer-passive / etc.) and attach flag bits to the candidate." sil_target: SIL2 asil_target: QM dal_target: DAL C iec_62304_class: NONE