Governance before execution.
Project-AI documents an execution-governed architecture so humans and AGI can eventually act as partnered stewards of Earth — accountable before any consequential action. Latency, enforcement, receipt, replay, and public-witness claims remain unverified until their promotion criteria and evidence bundles pass.
Start with claims, verification, and replay instead of the full manifesto scroll.
Use the quick tour, assurance brief, and catastrophic scenario framing.
Jump directly into the runtime model, kernel surfaces, and doctrine.
Go to the assurance brief and institutional contact path.
The wider practice of partnered guardianship — Design Bench, manifesto, doctrine.
- LivePublicly operational right now.
- SimulatedBrowser/demo execution only — not real enforcement.
- DocumentedArchitecture or behavior is described, but not exposed as a live public runtime.
- PlannedExplicit future work that is not yet exposed.
- Sign-inAvailable only behind authenticated or protected flows.
- LockedIntentionally restricted from public interaction.
The architectural bet, in three lines.
Deny by default
Contract requirement: identity, capability, policy, ethics, security, consistency, and audit must be satisfied before a governed effect is admitted.
Governance artifact pipeline
The TSCG-B codec exists. Compiler integration, signed packaging, and Code Store loading remain planned or unverified capabilities.
Three terminal verdicts
ALLOW · DENY · SAFE_HALT. The normal server path emits receipt-v1 bundles and the browser can verify supplied bundles locally; deployed key pairing, ledger provenance, and deterministic replay remain open.
OctoReflex. A documented pre-effect containment target.
OctoReflex is the documented engineering core. The constitution, nine gates, and Triumvirate are the why. The design calls for a deterministic evaluator on the effect path. Syscall-wide coverage, the latency bound, deterministic replay, and receipt coverage are assurance claims awaiting versioned execution evidence.
OctoReflex
Documented target: evaluate capability and policy before a governed effect. Complete syscall interposition and cross-host replay are not yet verified.
Policy packaging
The TSCG-B codec is implemented at the pinned source revision. Compiler integration, signing, and Code Store loading are separate planned or unverified capabilities.
Named refusal points
Ingress · STATE_REGISTER · Identity · Capability · Policy Binding · Cerberus · Galahad · Codex · Audit. Three verdicts: ALLOW · DENY · SAFE_HALT.
Planetary Continuity. Shared infrastructure for legitimate authority under recursive, cross-domain AI.
Written, signed, and published as a contribution to the commons — for humanity, and for the machine intelligences that will share this planet with us. The autonomy race is not won by whoever fields the fastest model; it is survived by whoever keeps accountable authority over machine speed. No institution, and no species, should hand the control plane of consequential action to something it cannot inspect. Project-AI is an open governance substrate: the constitution, the keys, and the audit chain stay with whoever is answerable for the action — and stay legible to everyone the action reaches.
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, we are betting the biosphere on hope. The durable advantage — for people and for AI — is governance enforced before execution. Project-AI specifies that target; this portal will describe it as verified only after the enforcement, receipt, and reviewer proof chain passes. Decision integrity is how two kinds of intelligence stay accountable to each other, and to the planet they both depend on.
"How do we preserve legitimate authority, decision integrity, and operational continuity when autonomous and cross-domain AI systems coordinate at machine speed?"
Project-AI's documented design organizes three engineering primitives — OctoReflex, the planned Code Store integration, and the nine execution gates — map directly onto the three things that must survive: authority, decision integrity, and operational continuity.
Authority artifact target.
The contract calls for a signed, hash-anchored constitution. TSCG-B signing and Code Store loading are not evidenced in the current pinned implementation.
Decisions should carry proof.
The contract maps doctrine to named gates and the terminal verdicts ALLOW · DENY · SAFE_HALT. Complete gate coverage and cryptographic verification remain open promotion criteria.
Machine-speed governance target.
L_reflex < L_cog is a deployment-specific assurance target. The portal does not treat it as universal until latency and bypass evidence are pinned and published.
Built to plug into critical command architectures, multi-party environments, and assured-autonomy programs — civil, scientific, or defensive — without surrendering the audit trail. The contract requires governance on the effect path and identity, capability, verdict, and receipt evidence for governed actions. End-to-end enforcement and ledger provenance are not yet verified.
A separately governed assurance layer designed to complement model-alignment methods in critical command architectures, multi-party environments, and assured-autonomy programs. This is an architectural position, not a benchmark result.
Signed receipt path
The normal server path hash-links and Ed25519-signs receipts. External timestamping is absent, and the current write boundary does not prove every stored row used that path.
Deterministic replay target
The receipt contract, canonical bytes, fixtures, and clean-clone harness are published in this source. Versioned cross-host replay and deployed-run evidence remain unverified.
Asymmetric security design
The architecture combines identity, capability tokens, and governed storage. Its bypass resistance has not yet been established by an end-to-end threat harness.
Effect-path containment
The design targets evaluation before governed effects. Syscall-wide coverage and the reflex latency bound await pinned runtime evidence.
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.
Cross-domain continuity target
The contract requires identity, authority, and chain-of-custody to travel with governed actions. Cross-domain proof remains a planned assurance exercise.
Inspect the kernel from any angle.
- LivePublicly operational right now.
- SimulatedBrowser/demo execution only — not real enforcement.
- DocumentedArchitecture or behavior is described, but not exposed as a live public runtime.
- PlannedExplicit future work that is not yet exposed.
- Sign-inAvailable only behind authenticated or protected flows.
- LockedIntentionally restricted from public interaction.