Governance before execution.
Project-AI is an execution-governed substrate. Containment is strictly faster than cognition. Nine gates, three principals, one signed audit chain.
The architectural bet, in three lines.
Deny by default
No meaningful action runs unless identity, capability, policy, ethics, security, consistency, and audit are satisfied first.
Compiled governance
Policy lives in a hash-anchored constitution. Validators load it from a signed Code Store at runtime — no advisory layer to ignore.
Three signed verdicts
ALLOW · DENY · SAFE_HALT. One append-only ledger. Deterministic replay from any receipt id.
OctoReflex. Containment at the syscall boundary, before execution reaches user space.
OctoReflex is the engineering core. The signed constitution, the nine gates, and the Triumvirate are the why. The reflex is a deterministic, side-effect-free pre-execution evaluator that runs at the kernel boundary — so containment is strictly faster than cognition by construction, not by policy. Every refusal cites a named gate. Every verdict produces a signed receipt.
OctoReflex
Syscall-level interposition. Deterministic, replayable, bit-for-bit. The action does not reach user space until the evaluator returns ALLOW.
Signed constitution
Policy compiled to TSCG-B binary, loaded at runtime from a hash-anchored Code Store. Tampering breaks the chain; there is no advisory layer to ignore.
Named refusal points
Ingress · STATE_REGISTER · Identity · Capability · Policy Binding · Cerberus · Galahad · Codex · Audit. Three verdicts: ALLOW · DENY · SAFE_HALT.
National Continuity. Sovereign American infrastructure for command legitimacy under recursive, cross-domain AI.
Designed, written, and signed in the United States. The autonomy race will not be won by whoever fields the fastest model — it will be won by whoever keeps lawful authority over machine speed. America cannot afford to outsource the control plane of its own command-and-control to foreign stacks, opaque foundation models, or vendors who answer to another flag. Project-AI is an American-built sovereign substrate: the constitution, the keys, and the audit chain stay under U.S. control.
The strategic case is simple. Adversaries are racing to deploy autonomous systems that act faster than any human can review. If the response is more probabilistic alignment bolted on after the fact, the United States is betting national authority on hope. The asymmetric American advantage is governance enforced before execution — deny-by-default, cryptographically anchored, and provable in a court, a hearing, or an after-action review. Decision integrity is national power.
"How do we preserve U.S. authority, decision integrity, and operational continuity when autonomous and cross-domain AI systems coordinate at machine speed?"
Project-AI's answer is mechanical, not aspirational. Three engineering primitives — OctoReflex, the signed Code Store, and the nine execution gates — map directly onto the three things that must survive: authority, decision integrity, and operational continuity.
Authority survives.
The signed, hash-anchored constitution is the source of truth for what any agent is permitted to do. Authority is cryptographic, not conversational — it cannot be re-prompted, social-engineered, or quietly overridden by a model update.
Every decision is provable.
Ten doctrine points compile to nine named execution gates with a finite verdict set: ALLOW · DENY · SAFE_HALT. Each verdict cites the gate that produced it and the constitution hash it ran against. Decisions are reviewable mechanically, not narratively.
Operations stay governed at machine speed.
Syscall-level interposition evaluates before execution reaches user space. Containment is strictly faster than cognition — by construction. Cross-domain coordination cannot outrun the audit trail, because the audit trail is on the critical path.
Built to plug into JADC2-style command architectures, mission partner environments, and assured-autonomy programs without surrendering the audit trail. Governance is enforced at the kernel boundary — not asserted after the fact by a dashboard. Every meaningful action carries an identity, a capability, a verdict, and a signed receipt that survives the action itself.
A deterministic substrate for assured autonomy: stronger assurance than probabilistic alignment, designed to plug into JADC2-style command architectures, mission partner environments, and assured-autonomy programs without surrendering the audit trail.
Append-only signed audit chain
Every verdict is hash-chained, Ed25519-signed, and timestamped against external TSA roots. Tampering breaks the chain; verification is a single command.
Deterministic replay
Any decision receipt replays bit-for-bit against the same constitution. Post-incident review, JAG review, and after-action become mechanical, not narrative.
Asymmetric security floor
Cryptographic identity, capability tokens, and a write-only governance vault. Attackers must forge signatures or break the kernel — not just persuade a model.
Syscall-level containment
Reflex containment evaluates at the syscall boundary before execution reaches user space. Containment is strictly faster than cognition — by construction.
Verifiable, not vibes-based
Ten doctrine points compile to nine execution gates and a finite verdict set: ALLOW · DENY · SAFE_HALT. Reasoning about safety becomes reasoning about code paths.
Command legitimacy preserved
Identity, authority, and chain-of-custody travel with every action across domains. The signed ledger is the record of record when speed exceeds human review.
Inspect the kernel from any angle.
- Livepublicly operational right now.
- Simulatedbrowser/demo execution only — not real enforcement.
- Documentedarchitecture described in repo / papers, no public runtime.
- Plannednot yet exposed.
- Sign-in requiredvisible to signed-in users only.
- Lockedadmin-only surface.