Receipts Beat Logs for Agent Audit Trails

Logs are useful for debugging, but agent governance needs stronger evidence than a stream of messages. When an AI agent requests a tool call, the important question is not only what happened. It is what was proposed, what policy decided, whether a side effect dispatched, and whether the evidence still verifies later.

HELM AI Kernel emits signed receipts for execution-boundary decisions. The local proof demo creates a signed DENY receipt, verifies it, then submits a flipped-verdict copy and confirms the tamper attempt fails verification.

Receipt Evidence Path

flowchart LR
    Request["Tool-call request"] --> Decision["ALLOW, DENY, or ESCALATE"]
    Decision --> Receipt["Signed receipt"]
    Receipt --> EvidencePack["EvidencePack"]
    EvidencePack --> Verify["Offline verifier"]
    Receipt --> Tamper["Tampered copy"]
    Tamper --> Reject["Verification failure"]

flowchart LR
    Request["Tool-call request"] --> Decision["ALLOW, DENY, or ESCALATE"]
    Decision --> Receipt["Signed receipt"]
    Receipt --> EvidencePack["EvidencePack"]
    EvidencePack --> Verify["Offline verifier"]
    Receipt --> Tamper["Tampered copy"]
    Tamper --> Reject["Verification failure"]

A useful agent receipt should make these checks boring:

• decision: ALLOW, DENY, or ESCALATE
• action identity and policy identity
• side-effect dispatch state
• receipt hash and signature state
• offline verification result
• tamper failure when the decision is modified

Run the proof path locally:

git clone https://github.com/Mindburn-Labs/helm-ai-kernel.git
cd helm-ai-kernel
make build
bash scripts/launch/demo-proof.sh

The demo uses localhost fixtures and sample policy data only.

Source Truth

• Verification
• Execution security model
• Proof launch demo
• Receipt schemas (protected staff doc)