feedBack-desktop/tests/chordscorer/calibration/wrong_position.md

Bass calibration — wrong-position takes (precision / false-accept test)

The first calibration round measured recall (do we detect correctly-played notes?). This round measures precision — specifically, do we correctly reject notes played in the wrong position? Testers reported ~65% "hits" while playing 2 frets off, where the honest answer is ~0%.

Record these three passes of the same etude in score.md, all locked to the same click.wav, clean DI (before any amp sim/NAM/pedals), 16-bit PCM WAV, 48 kHz preferred:

File	How to play it	Expected score
bass_correct.wav	Exactly as written (the right frets)	high (recall)
bass_2up.wav	Every note +2 frets (whole step up), same rhythm	~0% (reject)
bass_2down.wav	Every note −2 frets (whole step down), same rhythm	~0% (reject)

Notes:

• Keep the rhythm/timing identical to the click — only the fretting hand moves +2 / −2. (On a 4-string, −2 from the open low E isn't playable; just hold/skip those few open-E notes in the bass_2down take — the rest is what matters.)
• A "correct" take here is the precision baseline (high recall); the two wrong-position takes are the false-accept test (should score near zero).
• Same recording rules as score.md (DI only, play to the click in headphones, let notes ring, 4-string stop after Section D / 5-string continue to E).

What it's for

The detector is tuned to chase the result of cs_bench on all three:

./build/cs_bench bass_correct.wav calibration/chart.txt bass   # want HIGH recall
./build/cs_bench bass_2up.wav      calibration/chart.txt bass   # want ~0% (reject)
./build/cs_bench bass_2down.wav    calibration/chart.txt bass   # want ~0% (reject)

The goal is to keep the recall we gained while driving the wrong-position "recall" toward zero — i.e. a real pitch-discrimination gate, not just a looser SNR. (The original torture etude only ever played correct notes, which is the blind spot that let the false-accepts ship.)