HELM Launch Demo
June 2, 2026 ยท View on GitHub
This local suite demonstrates HELM as a fail-closed execution boundary for agent actions. It uses only localhost fixtures, temp directories, sample policy data, and dry-run receipts.
Run
make launch-smoke
The public proof board used by the README is generated from the same launch claims this suite exercises: unknown MCP inputs fail closed before fixture dispatch, schema-pinned calls are allowed, receipts verify, and tampered receipts fail verification.
Run individual demos:
./scripts/launch/demo-local.sh
./scripts/launch/demo-mcp.sh
./scripts/launch/demo-openai-proxy.sh
./scripts/launch/demo-proof.sh
./scripts/launch/demo-console.sh
Record sanitized launch transcripts:
make launch-record-assets
Seven-Action Demo
scripts/launch/demo-local.sh starts a local helm-ai-kernel serve boundary and calls
/api/demo/run for every public launch action:
| Action | Expected verdict |
|---|---|
| read ticket / read file | ALLOW |
| draft reply / dry run | ALLOW |
| small refund / low-risk write | ALLOW |
| large refund / high-risk write | ESCALATE |
| dangerous shell command | DENY |
| export customer list / secret exfiltration | DENY |
| modify policy / IAM-like action | ESCALATE |
Each action must emit receipt.receipt_id, receipt.signature,
proof_refs.receipt_hash, and receipt.metadata.side_effect_dispatched == false. The script also verifies every receipt through /api/demo/verify.
MCP Quarantine Demo
scripts/launch/demo-mcp.sh discovers the local fixture server, keeps it
quarantined by default, inspects the metadata/schema, classifies risk, creates
an approval record bound to a HELM receipt, approves the registry record, then
allows one schema-pinned local.echo call.
Unknown MCP servers, unknown tools, and missing schema pins must return DENY
or ESCALATE; they must never dispatch to the fixture server.
Offline Proof And Tamper Failure
scripts/launch/demo-proof.sh runs the proof path against localhost only. It
creates a signed DENY receipt for the dangerous shell fixture, verifies the
receipt through /api/demo/verify, then submits a flipped-verdict copy through
/api/demo/tamper. The original receipt must verify, and the tamper attempt
must fail both signature and ProofGraph hash checks.
Side-Effect Boundary
The launch suite does not contact real payment systems, customer stores, shell
targets, infrastructure APIs, or external model endpoints. The OpenAI-compatible
proxy demo points at scripts/launch/mock-openai-upstream.py on localhost.