THE DROSOPHILA CRITIC

The comparators are external regressions fitted on earlier synthetic runs (seeds 64–71). The fly is never trained or changed. Evaluation uses seeds 101–108; the human performance is not used for fitting. Reference evaluation repeats the calibration stimulus with new noise. All three comparators are retained. Residuals are approximation errors, not proof of a special connectome contribution.

Equal RMS does not mean equal stimulation. Human audio lasts 45.1 s versus 26.17 s and has 10 capped input frames. Stanza means exclude inter-stanza gaps. Human alignment is approximate; the boundary check does not validate it. Performance, duration and acoustics differ together, so this is not a causal effect of reader identity. The interpreter receives numerical response summaries, never the poem.

Targets use source-line intervals 2 and 17. The gain envelope transfers equal waveform energy between them, with 80 ms tapered edges. RMS is matched within PCM rounding; total injected drive is measured separately. No compression, pitch, timing or connectome changes. The amplitude edit has not been independently reviewed for naturalness.

Local windows and the response criterion were declared before 32 simulations with seeds 301–308. Criterion: temporal mean difference RMS exceeds across-seed variability RMS and the two four-run halves have cosine similarity above 0.5. This is a descriptive check, not statistical significance. Off-target means exclude both target windows and their 0.3-second extensions. The reading receives only response measurements, never poem text. No movement, feeling or preference is established.