Hall Security
A Rust-based observe-only research instrument for identity verification and security monitoring, with explicit safety boundaries.
observe-only security boundary
system flowEvidence moves toward review and audit. No path grants physical access.
The engineering problem
Hall Security (HallSight) is a research instrument, not an access-control product. It correlates camera observations, enrolled identity, and RFID presentation for authorized human review. It must never unlock or deny access to a building. That is a system invariant, not a feature flag.
Implementation evidence
The private repository separates hall-vision (Rust edge inference), hall-control (control, audit, persistence), hall-dashboard (operator UI), and shared contract types. In hall-vision/src/event_sink.rs, finalized events are written into a disk-backed outbox before being sent to the control service. Delivery happens separately from the inference loop, with retry/backoff and dead-letter handling.
This choice matters: an unavailable central server should not silently discard a finalized observation or freeze image inference while a request retries. A durable outbox alone does not prove zero event loss; that requires failure-injection and deployment qualification.
Other implementation and test evidence includes authenticated node communications, generated Rust-to-TypeScript contracts, observe-only policy tests, and a CI workflow covering Rust checks, dashboard verification, release tooling and documentation governance.
Why observation and action are separate
An experimental recognition model can produce insufficient, ambiguous, or incorrect evidence. The design leaves the physical door outside its authority and lets an operator review the evidence and uncertainty. It makes a narrower, defensible safety claim instead of implying biometric accuracy or liveness certification.
Verification status on October 10, 2026
The repository's October 9 release-blocker register still requires normal-node live tests, further GPU-host checks, calibration study, licensing decisions, and a qualified installable release. The latest GitHub Actions CI run visible in the audit (October 9) failed. Some tests and limited live demonstrations are documented, but they do not establish general operational accuracy.
What this project demonstrates: Rust architecture, separation of trust boundaries, durability patterns, defensive policy design, and an unusually explicit defect register.
What it does not demonstrate yet: a production-ready biometric security system, validated recognition accuracy in real deployment, or a qualified hardware installation.
Evidence inspected: hall-vision/src/event_sink.rs, .github/workflows/ci.yml, docs/qa/RELEASE_BLOCKERS.md, and recent Actions runs. Source is private.