# Local API boundary checklist

Proposed implementation review, not an enforced or tested configuration. Do not perform fault injection or external probes on a shared production service merely to fill this sheet.

- [ ] Record browser, backend and runtime host/network namespaces.
- [ ] Record exact runtime version and intended model identity/routing; localhost alone does not prove local inference.
- [ ] Inspect actual listener, host publishing, forwarding and intended access paths.
- [ ] Check intended and unintended reachability from the relevant controlled locations; record failures precisely, not as generic proof of isolation.
- [ ] For containers, distinguish host loopback, container loopback and private service addressing.
- [ ] Identify the actual authentication boundary and configured token requirement.
- [ ] Keep real credentials in the backend and outside source control and logs.
- [ ] Implement tenant/resource authorization separately from runtime authentication.
- [ ] Treat CORS as browser policy, not authentication or authorization.
- [ ] Implement input/output policy, admission limit, bounded queue and selected concurrency.
- [ ] Define per-operation timeout versus overall deadline; do not assume client disconnect cancels compute.
- [ ] Require protocol completion; inspect runtime errors even after HTTP200; classify partial answers as incomplete.
- [ ] Evaluate useful output separately from done:true or HTTP success.
- [ ] Use redacted diagnostics with deliberate access and retention controls.
- [ ] Complete the live matrix with synthetic data before accepting real user material.

## Evidence record
- Topology and location of each test:
- Runtime/API/model identity and local-routing evidence:
- Implemented access and resource controls:
- Actual observations: not_run
- Unknown outcomes and follow-up owner:

The companion client demonstrates bounded HTTP parsing only. It does not implement user identity, authentication gateway, authorization, queueing, deployment or accounting.
