Compare the computed and declared hash.
A canonical hash comparison checks record content under specified rules. It does not establish answer truth.
Trust begins with
a defined claim.
Read the boundaries of record integrity, key authentication, provenance and evidence support before drawing an assurance conclusion.
Start with the claim
Four questions. Four different checks.
Verification scope belongs beside the result. A public test key does not establish a production signer's real-world identity.
Verification scope
Match the check to the conclusion.
| Question | Check | What it does not establish |
|---|---|---|
| Has the record changed? | Recompute the canonical hash. | A match is about the specified record bytes. |
| Was the signed value authenticated by this key? | Verify the Ed25519 signature against a trusted public key. | The key must itself be trusted through a separate distribution process. |
| Does the source support the statement? | Review claim-level evidence and the disclosed signal method. | A signature or source reference does not prove support. |
| Was the computation actually performed? | Use an execution-evidence mechanism designed for that claim. | The current declared record is not execution proof. |
Current boundaries
Open work stays visible.
These statements describe the current evidence package. They are not hidden behind a general assurance label.
- 01
The current record chain uses HMAC, not a publicly verifiable public-key chain.
- 02
The Merkle root is unsigned and unpublished. Inclusion-proof implementation is not an externally witnessed log.
- 03
Complete omission detection requires further protocol work.
- 04
CONDUIT production Ed25519 integration remains open.
- 05
Migration of the raw UTF-8 audit.py request hash is separate from the closed signed-record defects.
- 06
Historical original source and extraction hashes cannot be recreated by a later backfill consistency check.
Removal is not erasure.
- Serving zeroing or removal plus a ledger event does not erase every stored artefact.
- Stored weight artefacts are not purged by serving-path removal.
- Source removal does not remove factual knowledge already present in the base model.
Separate tools are not an external audit.
- The Python and TypeScript verification tools are Nage-authored. They run separately from the service, but are not an independent organisational audit.
- The browser Record Lab uses a public synthetic test fixture.
- No regulatory certification or general compliance approval is asserted.
Inspect the evidence
Run the checks, then read their limits.
The public fixture lets you recompute its canonical hash and validate its Ed25519 signature without turning either result into a broader truth claim.