The book said "at bar 13 the singer sings these three notes." The computer looked at bar 13 of the song and did not find those notes. This page tries shifting the guess by one bar, two bars, etc, to see if the book meant a nearby bar. If none of the shifts make it right, the case gets sent to a human.
Offset strip legend: supported partial contradicted untestable
bar 13: B♭3 → C4 → C♯4 (card 2)
Offset probe ±10:
No clean shift found in ±10. Highest partial matches: offset -5 (2/3 beats match), offset +3 (2/3 beats match), offset +4 (2/3 beats match)
Bar 16: B3 → C4 → B♭3 → A3 (card 2)
Offset probe ±10:
No clean shift found in ±10. Highest partial matches: offset +7 (3/4 beats match), offset -8 (2/4 beats match), offset -7 (2/4 beats match)
Both Asleep vocal-sequence contradictions probe to manual_review, not a clean shift. The closest partial matches sit at offsets of five or more bars, which is too big a jump to accept blindly. The likelier root cause is that the walkthrough MIDI does not encode the same vocal content as the full-song MIDI the model-rebuild describes. P5 refuses to "fix" the ledger by inventing a shift the evidence does not clearly support.
~/Projects/_outputs/believeyoume-midi/autofix_offset.py~/Projects/_outputs/believeyoume-midi/asleep-in-the-trunk_autofix.jsonl~/Projects/_outputs/believeyoume-midi/asleep-in-the-trunk_claims_patched.jsonl (identical to original since 0 shifts accepted)~/Projects/_outputs/believeyoume-midi/asleep-in-the-trunk_verdicts.jsonl