---
title: "Signed Receipts for AI Agent Actions"
canonical: "https://helm.docs.mindburn.org/helm-ai-kernel/use-cases/llm-audit-receipts"
source: "helm-ai-kernel/docs/use-cases/llm-audit-receipts.md"
edit: "https://github.com/Mindburn-Labs/helm-ai-kernel/edit/main/docs/use-cases/llm-audit-receipts.md"
section: "use-cases"
access: "public"
sensitivity: "public"
last_reviewed: "2026-06-11"
checksum_sha256: "sha256:ac20a662d8cf27af76cd0b50e1d0780bc4678881acaf1de27dad606b0b7f1722"
build_timestamp: "2026-06-10T22:15:26.620Z"
---
# Signed Receipts for AI Agent Actions

Logs answer "what did we write down"; receipts answer "what actually happened, and can someone else check?" HELM AI Kernel signs every boundary decision with Ed25519 over a JCS (RFC 8785) canonical form, producing a tamper-evident record that includes the verdict, the policy hash, and the requesting identity.

Receipts roll up into EvidencePacks: content-addressed, SHA-256-hashed archives that bundle the decision chain for a run. Because canonicalization is deterministic and cross-platform, an auditor, customer, or counterparty can verify the pack offline without trusting the operator that produced it. Change one field and the signature breaks.

The norm this enables is simple: no receipt, no production. An agent action that cannot be replayed and verified did not happen as far as your audit trail is concerned.

## Audit Receipt Chain

```mermaid
flowchart LR
    Request["LLM action request"] --> Verdict["Boundary verdict"]
    Verdict --> Receipt["Signed receipt"]
    Receipt --> Hash["Receipt hash"]
    Hash --> EvidencePack["EvidencePack"]
    EvidencePack --> Auditor["Auditor verification"]
```

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

## Source Truth

- [Quickstart](/helm-ai-kernel/quickstart)
- [Execution security model](/helm-ai-kernel/execution-security-model)
- [MCP integration](/helm-ai-kernel/integrations/mcp)
- [Verification](/helm-ai-kernel/verification)