Episode Details

We correct a critical error from episode 12: the model wasn't actually mislabeling attack and block swaps—the greedy scoring algorithm was. The model produced correct labels in the correct order, but a greedy pairing system paired predictions with the smallest time differences first, leaving correct predictions orphaned and scored as errors. A deep dive into how a single algorithmic choice can completely misrepresent model performance. Timestamps: 00:00 - Episode 12 correction overview 01:15 - The attack/block swap details 04:30 - Greedy pairing algorithm explained (the runners analogy) 09:45 - How correct predictions became errors 18:00 - Why this matters for model evaluation --- Sources & further reading: • game-senser-fable — docs/podcast/LEARNING-CONTEXT.md §12 (§12.1 the resolution • result; §12.2 the ladder, the no-op parameter, the pose default; §12.3 first half • side-of-net and the silent-gap verdict; §12.4–12.5 for the cold open only) • game-senser-fable — docs/touch-arc/frontier-touches/EXP-18-RESULT.md (the headline table • per-type, the boundary gate, the far/near gap, the scorecard, cost) • game-senser-fable — docs/touch-arc/frontier-touches/ADOPTION-HIGH-RESOLUTION.md (Joey's • ruling, the scope limit, the re-pricing, the four-things-that-didn't-work summary) • game-senser-fable — docs/touch-arc/frontier-touches/EXP-17b-DECOMPOSITION.md (the ladder • the cancelling effects, the tolerance split, ace phantoms, the EXP-13 window-starvation • read) • game-senser-fable — docs/touch-arc/frontier-touches/EXP-17-RECEIVE-PROBE-RESULT.md (the • medium no-op, byte-identical token counts, the per-frame table, the deviation notice) • game-senser-fable — docs/touch-arc/frontier-touches/EXP-15-SIDE-OF-NET.md (the side • hypothesis, the 6-of-6 court-block sign flip, the two serve frames) • game-senser-fable — docs/touch-arc/frontier-touches/EXP-21-MATCHER-CROSSING.md (the • crossing mechanism, the three-matcher tables, block→attack to zero) • game-senser-fable — docs/touch-arc/frontier-touches/EXP-20-ATTACK-BLOCK.md (merge = 0 at • every window, both resolutions) • guist-bot — data/series/ml-volleyball/ERRATA.md (the two ep-12 corrections this episode • carries, and the ep-13 unknown it pays off) This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.