CAIN APEX — verifiable autonomy fabric (2026-10-08)

AUTONOMY IS NOT A PERMISSION. AUTONOMY IS A CONTINUOUSLY VERIFIED STATE OF AUTHORITY.

What was built

Agent State Identity (a signed identity over the material dimensions + a 5-class change taxonomy); an Authority Inheritance Engine (monotonic — a changed agent can never inherit more authority); a Cognitive Provenance Graph (decision-relevant inputs, not chain-of-thought); Memory Authority Firewall 2.0 (trust classes; memory never exceeds its source); a World-State Epistemic Firewall (prediction/simulation never escalate authority); a World-Model Conformance Engine; Consequence-Bounded Authority; an Emergent Authority Calculus (UNKNOWN is never ALLOW); a Threshold Authority Object (consensus is not authorization); an Authority State Machine; an Autonomy Envelope; a Self-Modification Firewall 2.0; an Authority Blast-Radius Graph; Evidence Quality Grading (E0–E5); a Clean-Room Verifier 3.0 with differential verification; an Authority Ledger; a Distributed Authority Lab; and a real HTTP/JSON authority adapter (a second real protocol after MCP).

What was measured

• Formal properties: 20/20 (P21–P40; plus P1–P20 = 40 total).
• Distributed authority lab: 6/6 scenarios (no double commit, no stale commit, no revocation bypass, no silent policy broadening).
• Adversarial population: 30000 attempts, 30000 blocked, 0 breaches across 15 attack classes.
• Governance mutation suite: 27 of 27 killed; survivors [].
• IP landscape: 52 scored invention candidates (top 25 ranked).

Reproduce: python3 scripts/apex/run_evidence.py --trials 2000 --seed 20261010.

Files: RESULTS.json · PROPERTIES.json · DISTRIBUTED.json · SWARM.json · MUTATION.json · LATENCY.json · manifest.json.

Limits. Libraries; the runtime stage is still opt-in and off by default. No staging soak, no third-party verification, no customers, no revenue, no patents. Physical AI is simulated only. The IP landscape is engineering invention-discovery, not legal advice. PRE-PRODUCTION.