With fixed models, excluding a scan with a repeated trace reduced depth gain from 21.5% to 5.4% in seven mouse retinal drug records

We tested whether axon depth adds drug response shape prediction beyond a measured local light response. We fit two ridge models on 4,257 bipolar release regions from 19 control-only mouse records. All drug-coded mice were excluded from fitting. The primary paired test had 802 regions, 13 scans and seven mouse records with four stored drug codes. Added smooth depth terms reduced drug waveform error by 21.49%. The descriptive 95% mouse resample range was -7.43% to +29.86%, with predictions fixed. Native checks found one TPMPA-coded scan with 64 identical stored local drug vectors. Each vector had 2,048 samples. Every local quality index was 1. Fifty-one entered the test. This scan gave 89.41% of the net added-depth score gain under the original weights. Excluding only that scan, without refitting, reduced gain to 5.44%, with a range of -9.30% to +12.04%. Every other single scan exclusion retained at least 21.17% gain. A later broad-layer model gave 4.34% gain after the same exclusion, with a descriptive range of 0.45% to 7.40%; this result is exploratory. Test depths were only the two ChAT bands. The source of the repeated vector is unknown. No trace repair or receptor mechanism is established. The scan audit was added after waveform inspection. Separate trace normalization, uncertain common row pairing and small treatment groups limit interpretation. The package preserves native vectors, source hashes, code, compact data, model states, all results, four figures and repeat steps. Full source and offline repeats check the published files.

retina · bipolar cells · structure and function · data audit · pharmacology · axon depth · prediction transfer · glutamate imaging