Architecture
Base informatique Trusted en 8 packages de HELM, protocole d'exécution et priorité des politiques.
Architecture
One thing to remember: HELM's entire trusted core is 8 Go packages. CI enforces that nothing else touches the governance boundary.
HELM is a fail-closed execution authority for AI agents. It sits between intent and effect — every tool call, sandbox execution, and self-extension passes through a governance boundary that produces signed, causal, deterministic proof.
Design invariants
| Invariant | Mechanism |
|---|---|
| Fail-closed | Unknown tools, invalid args, drifted outputs → DENY |
| Deterministic | JCS (RFC 8785), SHA-256, Ed25519, Lamport clocks |
| Auditable | Every decision → ProofGraph node. EvidencePacks verifiable offline |
Verified Planning Loop (VPL)
The canonical execution protocol for every tool call:
Request → Guardian (PEP)
If DENY → DenialReceipt → ProofGraph →
403If ALLOW → SafeExecutor → Tool Driver → Canonicalize → Receipt → ProofGraph → Checkpoint (Proof Condensation)
Step by step
- Model generates a tool call proposal
- CPI (Canonical Policy Index) validates the proposal schema
- PEP evaluates policy rules
- DENY → signed DenialReceipt appended to ProofGraph →
403 - ALLOW → SafeExecutor runs the tool driver
- Effect is JCS-canonicalized and hashed into a Receipt
- Receipt linked to the ProofGraph (causal DAG, Lamport-ordered)
- Periodic Proof Condensation creates Merkle root checkpoints
Policy precedence
Policies compose in strict priority order. Lower layers cannot override higher ones:
| Priority | Layer | Role |
|---|---|---|
| P0 | Ceilings | Hard limits — cannot be overridden |
| P1 | Policy Bundles | Organizational governance (signed) |
| P2 | Overlays | Runtime, per-session, per-agent |
| CPI | Validator | Deterministic policy stack validation |
| PEP | Executor | Guardian enforces, Executor runs |
- P0 — absolute ceilings (budget maximums, forbidden effect types)
- P1 — signed policy bundles (organizational governance)
- P2 — runtime overlays (session-scoped, can only narrow P1)
- CPI — validates composed policy stack for internal consistency
- PEP — Guardian applies resolved policy, produces signed DecisionRecord
Trusted Computing Base (TCB)
The kernel TCB is 8 packages — the minimal trusted core. CI enforces forbidden-import gates to prevent scope creep.
| Package | Purpose |
|---|---|
contracts/ |
Canonical data structures (Decision, Effect, Receipt, Intent) |
crypto/ |
Ed25519 signing, JCS canonicalization |
guardian/ |
Policy Enforcement Point (PEP), PRG enforcement |
executor/ |
SafeExecutor with receipt generation |
proofgraph/ |
Cryptographic ProofGraph DAG |
trust/registry/ |
Event-sourced trust registry |
runtime/sandbox/ |
WASI sandbox (wazero, deny-by-default) |
receipts/ |
Receipt policy enforcement (fail-closed) |
Data contracts
| Contract | Fields |
|---|---|
| DecisionRecord | Verdict, ReasonCode, PolicyDecisionHash, Ed25519 signature, LamportClock |
| Effect | ToolName, EffectType, InputHash, OutputHash |
| AuthorizedExecutionIntent | DecisionID, Guardian signature, TTL |
| Receipt | EffectHash, OutputHash, ArgsHash, PrevReceiptHash, LamportClock, Ed25519 signature |
| EvidencePack | Receipts, MerkleRoot, ProofGraphHash, Ed25519 signature |
External interfaces
| Interface | Endpoint |
|---|---|
| Standalone proxy | POST /v1/chat/completions, GET /helm/receipts, GET /helm/proofgraph |
| Kernel API | POST /v1/chat/completions, GET /api/v1/proofgraph/sessions, GET /api/v1/proofgraph/sessions/{session_id}/receipts |
| MCP gateway | GET /mcp/v1/capabilities, POST /mcp/v1/execute |
| Health check | GET /healthz |
Deployment patterns
| Pattern | Description |
|---|---|
| Sidecar proxy | helm proxy on :9090 — single base_url change in your app |
| MCP server | helm mcp serve for MCP-native clients |
| Gateway | Full kernel/API surface on :8080 for multiple agents/services |
| In-process | Embedded as a Go library |
Next steps
| Goal | Guide |
|---|---|
| Get running in 5 minutes | Quickstart |
| Understand conformance levels | Conformance |
| Review the threat model | Trust Model |