# Annotated population differences v6

Retrospective discovery and held-out-noise validation on saved original FlyBrain runs; no new simulations or simulator changes. Discovery/calibration: v1 seeds 64–71. Validation: v2 seeds 101–108, withheld from population selection and gain fitting, but already examined for pooled project results. The performances repeat: validation is across simulator noise, not new speakers, stimuli or biological flies.

Anatomical eligibility before response inspection: nonempty annotated cell type, at least five total cells, at least three direct postsynaptic partners of injected JO-A/B neurons, and no injected neurons. Use whole-type archived counts; these groups may include cells outside the direct auditory targets. Record total and directly connected membership. Exclude unannotated cells and composite comma-separated labels, which do not identify a single type. No category is inferred from the spelling of a name. This limited screen does not cover every downstream pathway or small identified motor circuit.

Use complete native-clock 100 ms bins within five stanza windows from v5. Do not warp time. Subtract exact-duration same-seed silence. Fit one nonnegative zero-intercept gain for each type and each of the three predeclared input-only comparators (instantaneous injection, causal 0.1/0.3 s smoothing) on discovery reference runs only. Hold gains fixed for validation. Compare human-minus-reference stanza mean rates, and residual differences after each comparator. Human onset/offset boundaries vary independently by −0.2, 0, +0.2 s; only complete bins lying inside each interval count. Report actual analyzed bounds.

A discovery type/stanza candidate must have raw and all three residual contrasts passing abs(mean)>sample SD and >=7/8 signs agreeing, at all nine boundary shifts. Residual signs must agree with raw direction. Rank eligible candidates by their weakest absolute mean/SD over the residual models and boundary shifts, select at most five; break ties by type name then stanza. Freeze selected identities/windows/directions in discovery.json before loading validation responses. Report the full screen, including failures, and eligible/search counts.

On validation, apply the same tests to the same candidates with no reselection. A survivor must pass all checks with the discovery direction. These are conservative descriptive gates, not p-values or family-wise error control; discovery scans many pairs. Retain failed candidates and nulls. Baseline failures demonstrate limitations of these three approximations, not connectome-specific causality, memory, emotion or action. Functional interpretation requires documented evidence; otherwise describe only anatomical association and rate differences. Interpretation receives numerical evidence and population labels, never poem text or performer identity.

Publish extraction hashes, fixed coefficients, all discovery statistics, selected candidates, validation values and boundary checks, and a compact count archive for eligible types across both ensembles. Preserve existing results. This is one bounded screen; no extra runs or relaxed thresholds if it yields a null.