Loop's first portfolio/research release makes its central promise reproducible: the model may claim completion, but only the user-confirmed contract and independently re-executed evidence can accept a strict task.
Highlights
- one-command, zero-key
make demofrom a fresh clone - browser golden path from task publication through verified Receipt replay
- server-enforced authority, step/token budgets, verification reserve, and rabbit-hole controls
- isolated Docker/Kubernetes execution with destination-bound production egress
- replayable Receipts binding contract coverage, checks, artifacts, provenance, and ledger integrity
- 12-case real-provider evaluation manifest plus an explicitly non-model-quality deterministic smoke result
- honest threat model, residual risks, architecture, strategy, and changelog
Verified release gates
- 414 backend tests; 72.48% branch-aware coverage against a 70% CI floor
- TypeScript lint/typecheck, 30 JavaScript tests, and production build
- Playwright strict-contract task and Receipt replay
- dependency audits with no known vulnerabilities
- Redis restart, worker reclaim, fail-closed readiness, and revocation acceptance
- disposable Kubernetes deploy, task, Receipt, NetworkPolicy, failed rollout, and rollback
- packaged startup smoke on macOS, Windows, and Ubuntu
- published
0.1.0GHCR images for API, Web, sandbox, and provider gateway
No paid provider or subscription-backed model was invoked to produce the committed smoke result.