Episode Details

What happens when you stop trying to see the hands and start tracking the ball instead? The lab abandons explanation and embraces detection—building a five-link experimental chain to catch contact misses that fooled expensive AI models. Spoiler: a simple filter outperformed them both. 00:00 - Intro: From explanation to detection 02:15 - The core insight: Watch the ball, not the hands 04:30 - Building the experimental chain: Five linked premises 06:00 - Link 1: Is the ball even visible when it matters? 08:45 - Link 2: Detecting which rallies matter (with zero GPU calls) 11:20 - The signature pattern: When labels inherit errors 14:00 - Four links hold, one breaks—here's what replaces it --- Sources & further reading: • game-senser-fable — docs/podcast/LEARNING-CONTEXT.md §12 (§§12.7, 12.8, 12.9) • game-senser-fable — docs/touch-arc/frontier-touches/EXP-26-30-BALL-REPAIR.md (the chain • the free gate, the failed search and its bimodal errors, the deliberately-unrun variant, the • saturated pre-declared test, and its own correction footer) • game-senser-fable — docs/touch-arc/frontier-touches/EXP-31-AUDIO-GATED-RESULT.md (the • signal swap, the held-out 6/9 at the boundary, the chance floor and shifted null, the repair • under both matchers, the per-type deltas, the six-second reproduction, the not-adopted • status, and the two corrections it makes to the record) • game-senser-fable — docs/touch-arc/frontier-touches/EXP-22-POSE-CORPUS.md (audio 6/7 vs • pose 3/7 at matched density; the lane dies, the corpus survives) • game-senser-fable — src/exp31audiogated.py / src/exp31_verify.py (the search and the • checks run against its own headline before it was quoted) This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.