Cornerman rule debug viewer

A static, dependency-free web app for inspecting BlazePose Drive caches against the Cornerman rules engine, frame by frame. BlazePose is the only engine it reads: Apple Vision, glove-wrist, YOLO, RTMPose, MoveNet and YOLO11 caches (Drive archive/) are skipped, and the engine-era comparison lenses live in cornerman-archive/legacy-pose/.

Live: https://cornerman-ai.github.io/debug-viewer/

Also here: the pipeline viewer (https://cornerman-ai.github.io/debug-viewer/pipeline/) — one full-pipeline test video with its punches, impact frames, defensive moves and facing angle on one timeline, read from Drive Ambo/data/pipeline_tests/. A separate page with its own README (pipeline/README.md).

What it does

Pick the cache folder once per session, then for each round pick just the video — the viewer auto-matches the cache files by filename. (A manual multi-file picker is still available as a fallback.)

The viewer renders the video with a skeleton overlay and a side-panel "lens" that re-paints the overlay and surfaces metrics for one rule at a time. Add a new lens by dropping a file in the right js/lenses/ subfolder (rules / models / research / inspect) and registering it in js/lenses/registry.js — see js/lenses/README.md.

Files stay on the user's machine — the page uses the browser File API, nothing is uploaded.

Workflow

1. Cache folder — pick Ambo/data/skeleton_data/blazepose/. The viewer indexes every `_blazepose_r.npy +
2. Video — pick an .mp4. If its basename matches an indexed video, the viewer auto-loads the only round, or shows a round dropdown when multiple rounds exist.
3. Lens — switch the right-side dropdown to "Guard drop" (etc).

Input format

BlazePose Drive cache, one set per round:

• <stem>_blazepose_r<N>.npy — float32, shape (N, 33, 8): per joint per frame x, y, z, x_world_m, y_world_m, z_world_m, visibility, presence. x, y are normalised to [0, 1] and de-normalised to pixels at load time using the video's natural dimensions. Lenses read it remapped to COCO-17 (state.pose, conf = visibility); the full 33 joints ride along as state.blaze33.
• <stem>_blazepose_r<N>_meta.json — at minimum { fps, layout: "blazepose33" }. Any other layout is refused.
• <stem>_blazepose_r<N>_pts.npy (optional) — each frame's source-video time; <stem>_blazepose_r<N>_punches.json (optional) — ST-GCN detections.

Pick the .npy and _meta.json together via multi-select in the pose picker.

Lenses

The lens dropdown in the app is the authoritative list — one module per lens in js/lenses/ (foldered by kind: rules / models / research / inspect), registered in js/lenses/registry.js.

Running locally

python3 -m http.server 8765
# open http://localhost:8765

Adding a new lens

One module in js/lenses/ (rules / models / research / inspect), one import + one entry in js/lenses/registry.js. The contract, what state holds, the DOM slots, the frame/time convention and the gotchas are in js/lenses/README.md — read that before writing one.