A-LIST exercise extractor (Core ML)

• Inputs: input_ids int32 (1,128) WordPiece/bert-base-uncased; attention_mask int32 (1,128); parsed float32 (1,10) regex features (see the app's ExerciseFeatureParser for the exact order and transforms).
• Output: float32 (1,8) — [log1p(duration_min), log1p(calories), heart_rate_bpm, intensity_1_10, steps_gate_logit, log1p(steps), distance_gate_logit, log1p(distance_km)].

Versions

• v2/ (current, 2026-08-23): identical graph retrained on the sol-relabeled dataset (exercise-text-unified:v0 — one uniform frontier teacher, calorie CV 1.6% vs the old labels' ~11%). On the frozen held-out-user sol eval (1,127 entries) it beats the v1 artifact on every field: duration 68.5% vs 62.7%, calories 58.0% vs 51.5%, heart rate ±10bpm 76.0% vs 67.6%, intensity ±1 81.7% vs 76.0%, steps gate 94.9% vs 88.6%, distance gate 94.6% vs 90.8%. Same I/O contract and footprint.
• Root files (v1, 2026-08-16): distilled from ~14.5k historical production extractions (mixed teacher generations). Kept resolvable for rollback; the app's ExerciseExtractorDescriptor.v1() still pins these bytes.
• golden_cases.json holds 10 input→output cases generated from the v1 artifact — the app uses them to pin the decode rules (weights-independent), so they remain valid across versions.