WHAT HAPPENED WHEN THOSE 60 SENTENCES WERE READ Recorded 23 August 2026. One speaker, one Mac. 'Mac heard right' = how many hard words survived plain recognition. 'After repair' = how many survived once the fix-up layer ran. The fix-up layer uses no AI model and no internet. Section lines hard words Mac heard after repair ------------------------------------------------------------------- Ordinary sentences 15 - - - One hard word each 15 15 4 11 Traps 12 - - - Split words 8 9 1 8 Spoken punctuation 5 - - - Long passages 5 15 5 9 ------------------------------------------------------------------- TOTAL 60 39 10 28 10 of 39 hard words survived recognition on its own. 28 of 39 survived after the fix-up layer. =================================================================== WHAT IT GOT WRONG =================================================================== One sentence was damaged by the fix-up layer itself: read: "He clawed his way back in." came out: "He Claude his way back in." A hand-taught rule (clawed -> Claude) fired on ordinary English. Taught rules are deliberately allowed to rewrite real words; that is the trade-off, and this is what it costs. Two other flagged words were the Mac mis-hearing the speaker, not the fix-up layer changing anything: "The code review took longer" heard as "The court review took longer" "it was a blood clot" heard as "it was a black clot" =================================================================== LIMITS, PLAINLY =================================================================== One speaker. One machine. One sitting. Read once, not repeated. No word-error-rate figure is claimed; this counts whether specific words survived, which is a narrower question. The audio is not included. Raw per-group records are in results-raw.jsonl next to this file.