ARC Open Governance

Examine

Do not take our word for any of it

This page is the examination, in the order an examiner runs it. Every step names what to run, what result to expect, and what that result does and does not prove.

A foundation that asks the world to build on new infrastructure owes the world the means to check it. The founding release is built for that posture. Its files are bound in a signed manifest, its history is a signed ledger, its claims are mapped to their evidence in one table, and its verifiers ship beside the machinery they verify. What follows is the whole path, with nothing held back that an examiner needs.

The examination path

Step one. Get the sealed artifact

The founding release archive stands at SHA-256 72d07526dc43eac89ccaca6772eaa16c48d8f66e1d623be6dbde593fab8c10dc, sealed against a manifest at SHA-256 cc00f45617177c10b99f8a30887d033d4e4605121351a2e01928eb84ff17a2f7. The repository is in private preview while the founding evidence gates complete, and a sealed examination copy is available on request through the contact section. An examination begins from the artifact itself, never from a description of it.

Step two. Check the archive against its manifest

sha256sum -c MANIFEST.sha256

The manifest binds 408 entries. The expected result is every line reporting OK and no line failing. A single altered byte anywhere in the tree fails this step loudly, which is the point of it.

Step three. Verify the signature

openssl dgst -sha256 -verify release-pubkey.pem -signature MANIFEST.sha256.sig MANIFEST.sha256

The expected result is Verified OK. The public key ships inside the tree, so the check needs nothing from us and no connection to us.

Step four. Verify the founding ledger

python3 ledger/verify-ledger.py

The expected result is the chain passing end to end, entries one through seventeen, every hash link and every signature checking against the entry before it. The entries themselves stand on the ledger page, with the digest of each file printed beside it.

Step five. Run the settlement demo

The Quickstart inside the release runs a governed settlement on your own machine, a proposal carried through settlement to a Statement. The screen recording of that run ships inside the tree, so the film itself is one more file the manifest binds, and the run you make can be compared against it.

Step six. Attack it

The tamper suite feeds the verifiers five corruptions, a corrupted payload, a forged signature, an altered ledger entry, a replay, and a wrong key. Every verifier must refuse every corruption, and the genuine control must still pass. The evaluation harness then maps every public claim to its evidence in one table, which today reads 16 demonstrated, 2 measured, 5 argued, and 5 not reachable at v0.1, with the five unreachable claims named at the same size as the sixteen.

What passing proves, and what it does not

Passing proves

The artifact in your hands is the artifact that was sealed, byte for byte. Its ledger is internally consistent and signed by the keys it names. Its claims table is honest about which claims rest on demonstration, which on measurement, and which are still argument. The machinery in the tree does what its evidence says it does.

Passing does not prove

It does not prove the fundamental claims of the research program. It does not prove production security beyond the properties the conformance matrix marks as tested. It does not prove independent federation, which waits on a node operated by someone outside the founding team. Reproduced is an independent team's act, and no page of ours can perform it for them.

The conformance snapshot

The release carries a security conformance matrix that rates each property in four words, tested, mitigated, partial, and out of scope, against the OWASP Top 10 for Agentic Applications and at the attack class level against MITRE ATLAS. Its four priorities today are these. Authenticated identity and settlement are tested within the reference implementation. Complete mediation and bypass resistance are partial, because paths outside the reference implementation remain unexercised. Cross action exposure conservation is partial, with concurrency at scale, retries, budget splitting, stale replicas, and interrupted transactions named as the unexercised classes. Independent Statements and recovery are partial, implemented and not yet adversarially exercised. The two attacks named for this system, authority laundering and evidence laundering, are both rated tested. The full matrix stands in SECURITY_CONFORMANCE.md inside the release.

The five hardest questions

These are the questions this program must answer under examination, stated as openly as we know how to state them. Each is an open target, and findings against any of them are welcome in the open.

One rule, end to end

Does the same rule survive the whole descent, from the mathematics through the calculus, the database, GMAP, KORE, the adapters, and the Statements they issue, with no step quietly weakening it.

The boundary in the deployed world

Can anything bypass the consequential boundary in a deployed environment, along a path the reference implementation does not exercise.

Disagreement between independent nodes

What happens when independent nodes disagree, under partitions, revoked keys, and competing witnesses, and whose record governs while they do.

Evidence that proves what it says

Does a Statement's evidence prove what the Statement claims. A signature proves a key signed information. It does not by itself prove the physical event occurred, and the distance between those two facts is where this question lives.

Strangers, safely

Can strangers reproduce and operate the release safely, with authentic source, exact builds, clean installs, recovery, signatures, rollback, and licensing all standing up without the founding team in the room.

Report a finding

Findings are contributions here, and the serious ones are the most valuable kind. The disclosure path stands in SECURITY.md in the release, and participation messages are acknowledged within two business days. Our own practice is on the record. When the steward's own agent failed on the consumer side this October, the failure was not quietly fixed. It was disclosed as entry eight of the founding ledger, where it remains.

Status, plainly. The founding release stands at NO-GO. Fifteen evidence gates must pass independent examination first, under scrutiny in three forms, scientific, adversarial engineering, and publication and provenance, with reviewers working independently before findings are compared. Blocking findings are corrected and retested, or the capability is excluded from the release and from its claims. Publication follows the gates, and the decision to publish is the founder's alone.