# Local amplitude emphasis v4 — declared before runs

Use the v1 synthetic PCM, unchanged in timing, samples outside two target passages, voice and pitch. Targets are existing source-line intervals 2 and 17. Text identifies display intervals only; no textual feature enters the encoder or interpretation.

Apply two complementary power-gain envelopes with 80 ms raised-cosine ramps inside each target interval. Let w1,w2 be their [0,1] windows and E1=sum(x²w1), E2=sum(x²w2). Transfer Q=0.45*min(E1,E2). Earlier emphasis has gain sqrt(1+Q*w1/E1−Q*w2/E2); later emphasis swaps the signs. This preserves total waveform energy before PCM16 rounding, without global gain, clipping, compression or changing any off-target sample. Gains are bounded between sqrt(0.55) and sqrt(1.45), approximately −2.60 to +1.61 dB. Save original PCM, processed WAVs, gains, RMS, peak and actual injected input. Quantization RMS tolerance 1e-6 absolute. Fail on clipping or changed off-target samples. The gain is smooth; no claim of human-audited naturalness or full prosodic emphasis. Equal waveform RMS does not imply equal frame-RMS drive; report actual dose/capping.

Conditions: unchanged reference, earlier emphasis, later emphasis, equal-duration silence. Eight new paired seeds 301–308; 32 full original FlyBrain simulations. Same initial reset, warmup, baseline, dt, weights and parameters; default 1 s tail. No optional stopping. Use the existing original RMS-to-JON encoder. No new receiver mechanism. Initial spikes must match across conditions before the first target. Raw spikes/counts and provenance retained; no new identical-input repeats because v1 and v3 already independently established exact repeatability.

Primary population: direct JON postsynaptic partners excluding injection cells. Descending neurons secondary; injected JO-A/B checks only. Comparisons: earlier−reference, later−reference, earlier−later. Primary windows: each target onset through target end+0.3 s, snapped to the 20 ms simulation clock, with complete 100 ms bins from that snapped start. Report all three comparisons and both windows; no selected peak search. In each window report average paired rate difference, per-seed values/SD/signs, temporal mean-trace RMS, pointwise seed-SD RMS, and split-half cosine (first four vs last four). Reuse the declared descriptive criterion: temporal mean RMS > variability RMS and half cosine >0.5. Not a p-value or biological inference.

Also report whole-audio mean paired rates and off-target mean rates excluding both target+0.3 s windows. Off-target counts may change through neural propagation even though waveform samples there are identical. Display stimulus-minus-same-seed-silence ensembles; compute paired differences from integer counts before rate conversion. Retain seed ranges as observed ranges, not confidence intervals. Show target-aligned navigation and actual audio/input/response clocks in the main listening interface, distinctly labeled as an ensemble experiment, not the chamber's earlier single-seed records.

Interpretation receives only response summaries, target numbers/times and criteria; no poem or voice identity. Explain whether local differences meet the criterion and whether output-related evidence exists. No pleasure, fear, preference, emotional intensity scale or bodily action claims. A response that follows changed amplitude can explain positive results; no history/memory claim follows.