# THIRE — Trusted Harness Integrity Runtime Enclave Read-only public console and machine-readable discovery. ## About THIRE governs the AI agents that operate critical infrastructure and essential facilities (power, water, transport, healthcare, commercial real estate and industrial facilities as the built-environment subset). It applies seven independent control layers to every agent action — bounded, observable, provable — and attests the result on a tamper-proof ledger. The newest attack surface in critical infrastructure is obedience without a check. THIRE is the check: the buyer is the executive accountable for the facility; the protected object is the agent; the consequence is physical. The product decisions, release posture, and governing gates are documented in the repo registers; this console is read-only. ## Surfaces - Landing page: `/` - Modules: `/modules.html` - Agents & characters: `/agents.html` - About Thire (persona + autobiography): `/about.html` - Console (read-only): `/console.html` - Bounty (coordinated disclosure): `/bounty/` - Machine-readable discovery: `/discovery/discovery.jsonld` - LLMs.txt index: `/llms.txt` - Intake contract: `/discovery/intake-contract.md` ## Three duties - ActiveDefense (guard): stand at the line; deny-by-default; fail-closed; red means stop. - ActiveAI (govern): keep the agent inside its job — scope, context integrity, human signature. - ActiveVuln (reason): reason about exposure and cost, on the record. ## Control layers 1. Scoped agent identity 2. Tool/command allow-list + declared actuator contracts 3. Deny-by-default egress + OT protocol allow-list 4. Vault-mediated credentials; no plaintext secrets 5. Memory & context integrity (prompt-injection control) 6. Human-in-the-loop authorization (physical/OT/financial writes) 7. Escape prevention & runtime attestation ## Governing gates - Draft-only; figures UNVERIFIED until a named artifact. - Money writes gated by `payment-gateway-greenlight`; payee must be assigned. - Red engagements require a signed scope contract; no standing targets. - Provisioning requires human approval (no-provision-without-approval). ## Intake Governed intake only: public submissions land HOLD + UNVERIFIED via the loopback intake (127.0.0.1, no external network). See `/discovery/intake-contract.md`.