# Temporal confirmation v2 — fixed before new simulations

Use new seeds 101–108 (not v1's 64–71), with no optional stopping. Preserve upstream FlyBrain, connectome weights, DT=20 ms and all neural parameters. Extend observation after sound to five seconds; this changes observation duration only. Warmup and initial baseline remain unchanged.

Conditions: v1 level-matched synthetic reference, human, redistributed pauses, reordered line chunks; a localized pause intervention on the synthetic PCM (move 180 ms of verified silence from after source line 8 to after line 4; preserve every speech sample and total duration); and an encoder-only reverse-frame control (reverse the reference injection sequence, preserving every injected value and integrated dose exactly). The last condition has no claimed oral performance or associated playback. It tests arrangement at the actual input boundary. Use same-seed zero input of the exact frame count and same tail for controls. Share a silence run only when clock, seed and duration match.

Prespecified populations: injected JO-A/B as input check, direct JON partners excluding injected cells as primary downstream measure, pooled descending neurons as secondary output measure; DNa02, DNp01, MDN, pIP10 are named exploratory output circuits, with no inferred action threshold or subjective feeling.

Primary temporal comparison: stimulus-minus-silence rate, then condition-minus-reference within each seed; complete non-overlapping 100 ms bins during the common audio interval. Report RMS of the mean paired trace, RMS of pointwise paired sample SD, and cosine similarity of the mean paired traces from seeds 101–104 versus 105–108. Descriptive confirmation rule: mean-trace RMS exceeds seed-SD RMS AND split-half cosine >0.5. This is a declared engineering criterion, not a p-value, confidence interval, biological validation or formal population inference. Retain nulls. Different durations mean the human elapsed-time comparison covers different passages; do not imply line-matched causal contrasts.

Secondary: full-audio mean rate difference and five-second mean tail difference; per-second tail bins, relative to each recording's own ending. Do not claim sustained persistence from one nonzero tail average.

Event analysis: fixed pre/post windows [-0.3,0), [0,0.3) seconds around nonblank synthetic/human line onsets, where windows fit within the observed audio. These are resumption/onset associations, not causal pathway latency estimates. Human alignments remain approximate. For localized intervention, separately report response around the first displaced resumption (line 6, after a blank source line), in each recording's own onset coordinates. This is a manipulation-specific diagnostic, not a reader-identity effect.

Interpretive episodes: select up to three non-overlapping 0.5-second windows ranked by absolute mean paired primary downstream change, within the common audio interval, with ≥1 second between starts. Report seed values/sign agreement and timestamps. This ranking is exploratory and selected on the confirmation ensemble; it is not independently confirmed event discovery. Supply only measured summary fields, population identity and temporal positions to a separate affect interpreter. Never supply poem text, waveform, performer name or textual content. Separate measured evidence, functional limits, and speculative affect reading. No fear/pleasure/attention claims; no actual movement inferred from pooled spikes.

Audit: save stimulus hashes, protocol hash, original simulator provenance, raw local spikes, complete population counts, compact public counts and interpretation input. Fail on changed cached inputs or mismatched control. Preserve v1 records; do not retroactively change its result.