A correction to episode 17: when a project scaled its grading system to hundreds of thousands of external data points, the metric swung by 627% while the model's actual behavior barely moved. The measurement tool itself was the trap—a deep dive into why statistics lie, what happens when denominators shift, and how a volleyball match pipeline exposed the flaws in scaling without thinking. The metric is the thing that moved, not the model. 0:00 - Correction: What episode 17 got wrong 2:45 - The 6x swing on a single match 6:30 - Match 540-4992: The first full pipeline run 10:15 - Serve/receive/attack/dig metrics break down 15:00 - Why the denominator matters more than the numerator --- Sources & further reading: • game-senser-fable — docs/podcast/LEARNING-CONTEXT.md §15 "The denominator that ate the ratio", written 2026-09-05 by the Fable session from inside the work, commit 8ce4003 (the first full match, the serve mechanism, the unseparable sets, the benchmark and the token-law confirmation, the interruption and the four hazards, the seven corrections, the two gate kills, the denominator finding on two skills, and the constraint that moved from budget to disk) • game-senser-fable — docs/podcast/HANDOFF-2026-09-05.md (the corrections owed to episodes 10 and 17, the pooled-serve "quote the pair" instruction, and the stale documents not to quote from) • game-senser-fable — docs/touch-arc/bulk/RETROSPECTIVE-2026-08-30.md (what the run answered, the serve diagnosis, the benchmark table, what broke, the seven corrections and the pattern in them, and what it says about a twenty-match campaign) • game-senser-fable — docs/touch-arc/bulk/BULK-1-544992-DECLARATION.md and BULK-1-544992-RESULT.md (declared before the paid run, including why this match and not the 720p one) • game-senser-fable — docs/touch-arc/bulk/EXP-41-RESULT.md, EXP-43-KAGGLE-PROBE.md, EXP-44-SYNC-VS-BATCH.md and SPEEDUP-SCOPE-2026-08-30.md (clip decimation killed, the gap scan kept, the Kaggle port killed on per-file latency, and batch measured at 4.0x slower for $1.29 less) • game-senser-fable — JOURNAL.md, entry of 2026-08-31 (the three-match pooled table, the two media-gate kills and their upstream diagnosis, the four infrastructure bugs found on first live use, and the pause) • game-senser-fable — src/launchrun.py, src/campaign.py, src/exp40bulk_boxscore.py and src/out/BENCHMARK-544992.json (the detached launcher, the cost authority, the bulk grader, and the transferable rates) This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.