Verification — proving the work

How it works — the pose layer. The verifier lazily injects TensorFlow.js + MoveNet Lightning (~1MB, UMD builds via CDN) the first time a verifier opens — never on app boot, never bundled. Every frame runs through MoveNet locally; the counter consumes only keypoints.

How it works — the counting state machine. Pure logic, unit-tested headlessly in count.ts: one tracked joint angle per exercise, two thresholds (hysteresis) plus a persistence window —

Anti-jitter rules, all enforced in code: keypoint confidence floor 0.35 (both sides averaged when both clear it); the 300ms debounce rejects twitchy half-reps; a tracking dropout longer than the debounce invalidates the phase (the person may have moved anywhere while untracked) — brief dropouts are tolerated. The Good-GYM angle-threshold approach (MIT), kept DOM-free so it can be reused by future verifiers (watch, IMU).

Confirming the counted set logs it straight into the match; the leaderboard's verified % rises with it.

How it works. Connects to the standard GATT Heart Rate Service (0x180D → measurement characteristic 0x2A37) — Polar H10, Garmin, Wahoo, anything that broadcasts it; no vendor API needed. Per-session it averages notifications into:

Settings (resting HR default 60, age default 35 → max 185) persist in localStorage["rwf.hr.v1"] and are clamped to sane ranges. Graceful by design: unsupported browser (iOS Safari), user cancel, and mid-session disconnect (partial averages are kept) all resolve cleanly — verified reps never block a workout.

The engine side: an entry carrying avgHrrPct plus a learned player baseline switches that entry's multiplier to the 70/30 measured-vs-declared blend (see Game Rules → tier handicap), with baseline drift ≤10% per update as anti-sandbagging.