v0.1.1: room-frame result, documents at the 09-29 state, site source

Tagged on the merge of PR #10, 2026-09-29 (PDT). Since v0.1.0, which predated PRs #4 and #5 and carried 124 tests, seven pull requests merged. Test suite at this tag: 193 passed, 1 skipped (the skip needs rclpy).

The room-frame result (PR #8)

The fine-tuned class split was scored on the 580 recorded floor-trials-1 frames under a plan committed and pushed before the run. R1 and R2 pass, carried by the counter camera: B 32 of 33, F 30 of 30, walking 0 of 90 read as lying. The floor camera read the head-first lie-down as standing in 30 of 30 frames and on its own would fail R1. One subject, one room, one lie-down per segment, near-duplicate frames; nobody fell. The model is the device-sized e65db0, not the one that passed F1 and F2 on public data. D1 stays open.

• Plan: docs/field-tests/2026-09-29-room-frames-plan.md. Results: docs/field-tests/2026-09-29-room-frames-results.md. Extract (counts, verdicts, hashes; no per-frame predictions): docs/field-tests/2026-09-29-room-frames/room-e65db0.json.
• Two adversarial reviews before the run and one of the results, with resolutions: docs/field-tests/2026-09-29-room-frames/reviews.md. 44 seeded mutations of the runner and scorer, each caught by a test: seeded_mutations.py and its output beside it.
• Runner jetson/rf_room_eval.py keeps predictions as returned and scores nothing; scorer tools/bag_analysis/score_room_frames.py gives a verdict only to a valid run (declared model, confidence, URL and server; pinned manifest and frame digests; 580 rows in manifest order; no request error; no unknown class name).
• Not verified at the run: the container's environment (no docker access without a password) and the network (no capture).

Also since v0.1.0

• PR #4: egress re-capture with METRICS_ENABLED=False: EG1 PASS at 12,987 bytes after the first capture's FAIL at 285,098 bytes.
• PR #5: DISABLE_VERSION_CHECK=True and YOLO_OFFLINE=True on the container command; two container-start captures (VC1 FAIL, then PASS); DR-11 records that the usage collector's aggregated record still leaves, by decision.
• PR #6: documents brought to the 09-28 state; three privacy statements scoped to what was captured; the Workflow README lists every flag.
• PR #7 and PR #9: website/, the source of the public page, with a claims table mapping each statement to a path:line; evidence block 9 carries the room result with the floor camera's failure beside it.
• PR #10: CITATION.cff 0.1.1.

Unchanged

The LIDAR detector, the replay harness and the decision records' earlier entries are as at v0.1.0. Research prototype measured on one subject in one room; not a medical device. Patent pending; the V-JEPA stage is not in this repository.

v0.1.0: LIDAR fall detector prototype, replay harness, pre-declared evaluations

Research prototype, measured on one subject in one room. Not a medical device, not cleared by any regulator, and not to be relied on to detect falls. Apache-2.0, patent pending (see NOTICE).

The tag points at 3adcd15, the merge of #3 on 2026-09-28 (17:25 PDT).

In this tag

• A LIDAR fall detector for a floor-level 2D LIDAR on a Jetson Orin Nano (ROS 2 Humble, with a ROS-free core).
• An offline replay harness that scores a detector change against the recorded bags. 124 tests pass and 1 is skipped (it needs ROS).
• Field results from 2026-09-25 to 2026-09-27, failures included: the LIDAR sees a person lying across or diagonal to its beam within a second and misses lying end-on (2 of 6 lie-downs); stock RF-DETR finds the person in those frames, and its box shape does not separate lying from standing (C3).
• A fine-tuned RF-DETR class split, under a plan committed before training. On a public test split of 73 images, lying scores precision 1.000 and recall 1.000, and 0 of 24 lying instances are read as upright (F1 and F2 pass). One training run, near-duplicate frames unchecked: an upper bound, and not a room result.
• Device fit on the Jetson, report only: a median 78.7 ms and 125.2 ms per frame for the two models (serial, one request in flight, cameras off).
• Co-load, one run: the LIDAR scan stream held its rate while a model served 580 frames (CL1 and CL2 pass).
• The first egress capture. It failed its pre-declared bar (EG1): the inference server was posting a record of every request once a minute. EG2 passed.
• 18 decision records.

Not in this tag

These landed on main afterwards, in #4 and #5:

• The egress re-capture with METRICS_ENABLED=False (EG1 pass at 12,987 bytes).
• Two container-start captures and the two switches they checked.
• The DR-11 decision on the usage record that still leaves while inferring.
• The test count of 130.

Limits

No room-frame accuracy exists for the fine-tuned models. A model download has never been captured. The device still runs the legacy detector; the fixes are default-off and tested offline. See Known limitations.