# Corresponding passages and amplitude baselines v5

Retrospective analysis of existing simulations; no new confirmation experiment and no changed fly model. Declare the analysis before calculating these results. Use the five stanza intervals (source lines 1–4, 6–9, 11–14, 16–19, 21–24), from first onset to final offset, on each performance's native clock. No audio stretching or temporal trace warping. Human timestamps are approximate editorial boundaries, originally intended for display. Stanza results are exploratory; repeat all summaries with each human boundary shifted independently by −0.2, 0, +0.2 seconds, clipped to audio duration. This is a sensitivity check, not validation of alignment. Inter-stanza gaps are excluded from stanza means.

Use full original v2 counts, seeds 101–108, each subtracting its own equal-duration same-seed silence. Pair stanza mean rates across performances by seed, despite their different elapsed-time intervals; pairing does not make these identical noise intervals. Primary group: direct JON partners excluding injected cells. Secondary: pooled descending neurons. Report each performance's duration, mean injected drive, integrated drive, excess spikes per neuron, rate and seed range. Report human-minus-synthetic rate difference with all seed values. Descriptive directional criterion: absolute mean > sample seed SD AND at least 7/8 seed signs agree. Boundary-robust means the criterion and direction survive all nine human boundary shifts. No p-values, no biological population inference, no ranking selected passages as independently confirmed discoveries.

Fit three deliberately small input-only comparators to v1 synthetic-reference seeds 64–71: instantaneous injection; a causal exponential filter with fixed time constant 0.1 s; one with fixed time constant 0.3 s. Start filters at zero. Each has a single nonnegative least-squares gain, zero intercept, fitted against the mean silence-subtracted primary downstream trace in complete 100 ms bins. There is no model selection. Neural weights remain frozen: these regressions are external analysis, never part of the fly pipeline. Evaluate all three on v2 reference and human ensembles. Reference evaluation holds out noise seeds but reuses the stimulus; human evaluation holds out both performance and seeds. Report RMSE, centered R² (may be negative), RMSE against a zero-response predictor, seed spread and residual paired stanza contrasts. Residual differences are errors of these narrow baselines, not proof of connectome-specific computation, memory, affect or behavior. A good fit says an amplitude-following approximation is adequate at this measurement scale, not that the connectome is dispensable in all respects.

Interpreter input: numerical passage IDs, rates, directions, boundary robustness and comparator scores only. No poem text, waveform or performer identity. Preserve qualified/null outcomes. UI poem and line timing remain display/navigation metadata, separate from the interpreter. Retain hashes of all source count, manifest, encoding and waveform artifacts, exact regression coefficients and predictions, per-seed outputs, and the interpreter input.