ship’s manual · guide
Verification and limits
A green check is useful only when its scope is explicit. 1Helm publishes several kinds of evidence and does not collapse them into one inflated reliability score.
Deterministic autonomy contract
npm run benchmark:autonomy verifies six named runtime invariants: substantive shipped playbooks, the narrow human-blocker gate, the bounded outcome gate, resident autonomy tools, wakeable recurring work, and audit-chain integrity for its executed fixture. The output is machine-readable against the public JSON schema.
It is not a live-model benchmark. A 6/6 report does not prove that every model will complete every task, that every provider will stay available, that unexecuted code paths are safe, or that the product has no security defects.
Behavioral and integration evidence
Native integration tests drive full resident turns, direct Skipper escalation, automatic hand-back, runtime recovery from operational hand-holding, durable follow-ups, computer lifecycle, files, memory, and crash recovery. Browser acceptance covers user-visible workflows. Release acceptance additionally exercises a real configured provider and a real scoped Gmail operation without publishing credentials.
External systems
Gmail, Photon, model providers, Cloudflare, Apple notarization, and operating systems have independent failure modes. Their behavior is claimed only to the level exercised in the current release. Text is the Photon contract while rich attachment fidelity remains under verification.
Audit-chain boundary
The audit chain detects changes within the retained chain. It is not a remotely witnessed transparency log, does not backfill historical rows, and cannot stop an administrator with database access from replacing the entire database.
How to judge a claim
- Contract check: the named invariant passed in the deterministic fixture.
- Integration check: the exercised product path passed in a controlled environment.
- Live acceptance: the named external provider or machine completed the stated operation.
- Not verified: no stronger claim is implied.