digital-drosophila/registry/catalog-20260929/evidence/phase4w-late-kc-feedback/audit/MECHANISM.md

Phase 4W mechanism summary

Descriptive audit; no learning, model parameter change, or candidate selection. Six previously seen odors, one fixed seed each; not independent population generalization.

neuron_idx	body_id	actual annotation type
59832	79362	KCg-s3
78836	104513	KCg-d
82312	109321	KC

neuron_idx	source class	fraction of positive late arrivals
59832	APL	0.0000%
59832	DAN	23.9674%
59832	KC	0.2979%
59832	MBON	0.4567%
59832	PN_olfactory	60.4124%
59832	other	14.8655%
78836	DAN	91.3692%
78836	KC	0.8757%
78836	MBON	0.0000%
78836	other	7.7551%
82312	APL	0.0000%
82312	DAN	0.6249%
82312	PN_olfactory	99.3751%
82312	other	0.0000%

neuron_idx	isolated replay	late spike count range
59832	base	15–17
59832	omit_APL	15–17
59832	omit_APL_overlay	143–144
59832	omit_DAN	0–0
59832	omit_KC	14–17
59832	omit_MBON	28–30
59832	omit_PN_olfactory	0–0
59832	omit_other	0–0
78836	base	39–40
78836	omit_APL	39–40
78836	omit_APL_overlay	39–40
78836	omit_DAN	0–0
78836	omit_KC	39–39
78836	omit_MBON	46–47
78836	omit_PN_olfactory	39–40
78836	omit_other	49–50
82312	base	60–61
82312	omit_APL	60–61
82312	omit_APL_overlay	115–115
82312	omit_DAN	59–59
82312	omit_KC	60–61
82312	omit_MBON	60–61
82312	omit_PN_olfactory	0–0
82312	omit_other	65–66

Replay holds all other recorded drive and the measured applied APL signal fixed. Removing a source class is a conditional single-cell test, not a closed-loop whole-brain intervention. Drive fractions refer to delivered positive synaptic input, not fractions of biological causality or of net g. Inhibition and APL are separately accounted. Ranked ancestry uses positive edges only; inhibitory influences are not claimed absent. Ranking is a heuristic; all path edges have real source events and arrival<=target-spike time. Float64 category shadow max absolute residual versus actual float32 g: 1.76084011e-05. This does not alter the exact CPU replay or actual queue checks.