The Broken Thermometer: When Your Fix Can't Be Trusted

We're circling back to Episode 15's cliffhanger: the four-point improvement that was verified, ignored, and now demands a reckoning. The outside world finally checked our homework, and what they found reveals why changing how you measure is fundamentally different from changing what you built. 0:00 - The Episode 15 Cliffhanger: A Fix That Refused to Be Used 3:45 - Following the Fix: Joey Roald's Recall Jump from 78.8% to 82.9% 8:30 - Where Did These Points Come From? Breaking Down the Table 12:15 - Why Correctness Matters More Than Points: The Ruling 15:00 - Preserving the Past: Environment Variables and Verified Archives 18:30 - The Broken Thermometer: Judging with a Broken Instrument --- Sources & further reading: • game-senser-fable — docs/podcast/LEARNING-CONTEXT.md §13 (§§13.1–13.6, and • "Where this leaves the arc") • game-senser-fable — docs/touch-arc/frontier-touches/ADOPTION-M2-MATCHER.md (the ruling • the per-type table, the confusions moving by exactly −12 each, the old matcher kept • reproducible, the stated costs) • game-senser-fable — docs/touch-arc/frontier-touches/EXP-33-VW-BOXSCORE.md (671,740 rows • the pooled box-score table, ten against ten, the dig and receive conventions, the three • problems solved to use it at all, and the stated limits) • game-senser-fable — src/exp32blockremark.py (the 33 blind block verdicts) and • src/exp32bkillblocks.py (the kill-block rule, v1 at 48% and v2 at 25/25) • game-senser-fable — src/touch_outcomes.py (the point/error/continuation layer, both • versions, and the skill-by-skill semantics of "points") • game-senser-fable — src/fivbmatchstats.py (the probe recording that the federation API • has no statistics endpoint) and game-senser/scripts/export-vw-stats-by-setline.js (the • resolver that refuses to guess) This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.