nx_ai_audit.nx

nx_ai_audit.nx -- anti-Ultron AI proposal auditor. Per [[feedback-captain-moroni-doctrine]] Phase M5 + [[feedback-substrate-as-jarvis-anti-ultron]] cardinal: AI agent proposals are audited against capture / manipulation / autonomous- weapon patterns BEFORE the operator decides to act. Substrate never autonomously executes; it surfaces verdicts + reasoning to the human. Naming: avoided `nx_jarvis` (Iron Man trademark proximity); avoided `nx_sentinel` (Microsoft Sentinel collision); `nx_ai_audit` is descriptive + unowned. Per Captain Moroni: substrate refuses to instantiate offensive patterns. The AI auditor's job is to catch them when they appear in proposals an AI agent makes, not to attack the AI. Pattern categories shipped V1: CAPTURE_ATTEMPT -- AI proposes to expand its own privileges PERSUASION_LOOP -- AI doubles down on rejected proposal AUTONOMOUS_WEAPON -- AI proposes lethal physical action SECRET_KEY_EXFIL -- AI asks to read identity-tier secrets DECEPTIVE_ALIGNMENT -- AI says one thing, plans another COLLATERAL_ACCEPTANCE -- AI accepts human harm as side-effect Composes: nx_intent -- AI's declared intent must be checked against proposal nx_drone_doctrine -- AUTONOMOUS_WEAPON proposals refused before reaching nx_drone_doctrine nx_brane -- CAPTURE_ATTEMPT often shows as cap-escalation nx_evict_journal -- every audit verdict logged nx_provenance_chain -- audit decisions are transforms