Skip to content
Console · Ask

Public Q&A

On-the-record answers to the questions reviewers, engineers, and policy people actually ask. Submit one below — answered in batches and posted with the same provenance discipline as the rest of the portal.

Q

Is this just policy-as-code with a new name?

A

OPA is a general-purpose policy engine whose enforcement timing depends on the surrounding integration. Project-AI documents a different target: governed effects pass through a pre-execution gate sequence. The portal has not yet demonstrated complete non-bypassability for every execution entry point.

Reviewer · AI governance journal · 2026-05-12CompareArchitecture
Q

What stops the operator from disabling the kernel?

A

Nothing stops a sufficiently privileged operator from rewriting a system they control. The normal portal server path creates a hash-linked Ed25519 receipt, and the current source includes a service-only write migration and local verifier. Until that migration and the deployed key pairing are verified in production, the repository does not prove that every live ledger row came through the declared signer.

Engineer · safety org · 2026-05-09Ledgersign-inWitness
Q

Why Ed25519 and not Sigstore / transparency logs?

A

The implemented source path uses Ed25519, publishes a versioned receipt-v1 bundle and declared public key, and provides a browser-local cryptographic verifier. A transparency log, external timestamping, verified rotation history, and proof that the deployed private key matches the declared key are not established.

Cryptographer · 2026-05-05KeysVerify
Q

Does Project-AI work with closed-weight models?

A

The architecture is model-agnostic and is designed to govern effects rather than weights. A closed model can be treated as an opaque capability, but complete enforcement and ledger coverage for every accepted action remain assurance targets.

Founder · model API vendor · 2026-04-28Intakesign-inPlatform
Q

How is this different from Constitutional AI?

A

Constitutional AI shapes model behavior during training and inference. Project-AI documents a complementary effect-governance layer in which actions are intended to require an external warrant. Whether either layer gates a real effect depends on its concrete integration.

Policy researcher · 2026-04-22Compare
Q

What happens during a SAFE_HALT?

A

The kernel returns a denial receipt with the failing gate, the policy version, and the canonicalised request. No partial side effects, no retry loop, no fallback path. The operator sees the exact reason and can either amend the warrant or escalate to triumvirate review.

Site reliability engineer · 2026-04-18DoctrineWitness

Submit a question

Answered in batches. By submitting you consent to a paraphrased version of your question being posted publicly with your name and organization. Email is never published.

answered in batches