Skip to content

Status and limitations

Hemkesh R edited this page Nov 16, 2026 · 2 revisions

Status and limitations — beta.1

Updated 17 November 2026 for v2.0.0-beta.1. The published release is an experimental beta. The feature status below follows the release notes and the session record. A beta version name does not change admission or establish production acceptance.

What beta.1 adds

  • An installer for a tester's own Jetson: pinned dependencies and model downloads, on-device builds, a device-artifacts record, checks, manual guarded launch, clean stop and uninstall.
  • LFM3-VL-2B as the default scene/chat model, using a pinned llama.cpp build and capped image-token profile.
  • Opt-in identity retention on a continuous track life, separate from the raw face votes.
  • The Settings dialog, browser-local appearance and operator-key-protected runtime controls.
  • Ask from records, follow-ups, reanalysis, automatic refresh, zoom and visible-room scan.
  • A conversation that scrolls inside Ask, an information popover and the r9 resize/following correction.
  • Focused regression fixes for held-frame expiry, chat cancellation/disable behavior, zoom ownership and explanations.

Evidence and remaining limits

Area Evidence available What remains unverified or unchanged
Core monitoring Guarded reference-device checks, including one real Telegram send; portable regression tests Not production-accepted; the tester profile has external notifications off
Dashboard and controls Browser tests with synthetic data; reference-device observations and operator feedback r9's resize fix was exercised in headless Chromium with list scroll anchoring disabled, not a new live/model session or a native browser without anchoring
Face recognition Enrollment checks and single-person observations; the maintainer reported satisfactory behavior on 11 November F2 has never passed; no false-acceptance data; formal accuracy and multiple-person testing remain pending
Identity retention Implemented with explicit track-life boundaries, contradiction rules and expiry; regression tests and operator feedback Opt-in validation mode; proposed policy values and wrong-identity retention rates are not validated
LFM3 with face and chat These lanes have run together on the reference device. One capped integration run passed execution, recognition-execution and memory checks That integration's overall reading did not pass because cancellation was missing; a separate cancellation check is separate evidence
Visual chat and records answers Deterministic records routing, bounded history and queries; browser/runtime tests; single-person operator feedback Chat is not admitted; model answer accuracy, follow-up/region/scan quality and combined formal validation remain pending
Memory Guarded observations exist; a 9 November observation completed, and an 11 November observation was stopped by the memory guard No general resource guarantee; the cause of the later pressure rise was not established; retained-frame overhead is not an acceptance measurement
Installer r5 completed a camera-free build rehearsal on one reference Jetson. r6 changes were tested offline; r7/r8 changed documents/version; r9 added the browser fix and passed offline archive checks No complete launch from this package, no complete real-build check with enrollment, no enrollment through the package and no second Jetson trial

The earlier Wiki statements “face recognition and the VLM have never run together” and “no device installation guide” are superseded. Integration observations and an installer now exist; their limits above still matter.

Installer platform and memory

The reference system is one Jetson Orin Nano 8 GB, L4T R36.4.7 with Ubuntu 24.04 userland and an extracted TensorRT Python binding. Stock JetPack 6 / Ubuntu 22.04 is unverified. Preflight labels an unverified platform, and continuing requires the documented opt-in.

The default admission is tied to the reference device's artifacts. --device-artifacts FILE lets an experimental run identify a tester's own verified builds; it is not resource validation. No procedure has yet validated that route's resource limits on another device.

One headless engine export finished in 187 seconds starting with 6.95 GB free, after two unsuccessful attempts. The chosen launch/build gates, the monitor's stop thresholds and observed memory usage are different quantities. The minimum memory needed and general build times are unmeasured.

Release checks

On the tagged commit 59be406, the Dashboard workflow and portable suite passed:

The skips require device-only tools, reference artifacts or optional packages absent from CI. CI is not camera, GPU, memory or recognition validation. The archive was built from 5a1238a; the tag adds documents only, with no packaged-file difference. The release notes document archive reproducibility and the r8-to-r9 comparison.

Data and access

  • Inference is local. Visual questions are answered by the local model; supported records questions use local records.
  • The gallery is sealed with a passphrase. Public issues must not contain credentials, control keys, face data or footage.
  • The dashboard is loopback-only and reached through SSH. It has no general multi-user login/roles system.
  • Runtime feature changes require the operator key. This is distinct from access to the dashboard itself.
  • Snapshots and Telegram are optional capabilities; the supplied tester launch profile disables them.

Deferred work

Multiple-person testing, formal face/answer-quality acceptance, a broader device installation trial, resource validation on other devices and longer reliability checks remain open. Additional cameras, clips and a full authenticated multi-user API are not part of this beta. PIR/LiDAR and other sensor/model experiments remain deferred.

This page updates documentation only; it changes no threshold, enrollment, admission or recorded result.