Manifest
A canonical index of every artifact in the pack, hashed so the contents cannot be swapped or trimmed.
Proof
An EvidencePack is a content-addressed, signed bundle of receipts and proof-graph edges. It is built to verify offline, so a reviewer or a customer can check what your AI agents did without trusting your console.
Hashed, not stored. Signed, not logged. Verified by anyone, not just you.
What it is
When an AI agent takes a sequence of consequential actions, the evidence is scattered across logs and consoles. An EvidencePack collects the signed receipts, binds them with the proof-graph edges that show how they relate, and addresses the whole set by a content hash. The pack is the proof, not a pointer to it.
A canonical index of every artifact in the pack, hashed so the contents cannot be swapped or trimmed.
The signed verdicts and bound effects for each action, each one verifiable on its own.
The causal links between intents, verdicts, and effects, so the order of events is part of the record.
The policy hashes each action was checked against, so a reviewer can see the rules that applied.
A single hash over the canonical manifest, so the whole pack is addressed by its contents.
Everything needed to check the pack travels with it. No live HELM, no dashboard, no trust in the source.
Why offline matters
The people who most need the proof are the ones outside your systems: a security reviewer, an auditor, a customer doing diligence. An EvidencePack is built for them.
How to verify
Verification is a local check. Anyone with the pack and the public key can run it.
Step 1
Receive the EvidencePack as a file. It carries everything needed to check it.
Step 2
Hash the canonical manifest and compare it to the pack’s content hash.
Step 3
Verify each receipt signature with the public key. No live service required.
Step 4
See the ALLOW, DENY, or ESCALATE decision and the effect it was bound to.
Questions
A content-addressed, signed bundle that holds a manifest, the receipts for a set of actions, and the proof-graph edges between them. It is built so it can be verified outside any dashboard, by someone who does not run your systems.
A record you can only read inside a vendor console is only as trustworthy as the console. An EvidencePack is hashed and signed, so a reviewer, an auditor, or a customer can verify it on their own machine without trusting the source.
A receipt proves one action. An EvidencePack collects the receipts for a sequence, adds the proof-graph edges that show how they relate, and binds the whole set under one content hash.
Recompute the manifest hash, compare it to the pack’s content hash, then check each receipt signature with the public key. If the hash matches and the signatures verify, the pack has not been altered.
Keep reading
Terms
A small bundle of records used to verify one event or review path.
Use for replayable evidence slices.A record chain that helps replay and check what happened.
Use for HELM proof records and replay paths.HELM lets the action run.
Use as a canonical verdict.HELM blocks the action.
Use as a canonical verdict.HELM stops and asks for more facts, policy, or human approval.
Use as the canonical non-dispatch path for missing facts, policy hold, or approval.See a signed EvidencePack and verify it yourself, the way a reviewer would.