Experiments · E6

Was the old importer keeping all of the fly brain's connections?

No. Its five-contact cut-off discarded 76% of them; the full brain has 25,582,938 connections, extracted in 11.5 seconds.

In the log: The whole MaleCNS asset, and what the importer was discarding

recordedDate 2026-09-12, as written in the loggenerator · fly brain0 predictions · 1 result paragraphEXPERIMENTS.md lines 174–204

Checked · 2 discrepancies

Files checked (5)

data/body-annotations-male-cns-v1.0-minconf-0.5.feather
data/body-neurotransmitters-male-cns-v1.0.feather
data/malecns_full.json
data/malecns_circuit.extraction.json
forager/brain/connectome.py

Numbers confirmed (4)

claim: 166,700 nodes = every body with an assigned superclass; source: annotations feather: 166,700 bodies with non-null superclass; ok: true
claim: Transmitters: ACh 103,720, glutamate 29,302, GABA 22,069, histamine 7,891, DA/OA/5-HT 541; source: neurotransmitters feather, consensus_nt over those bodies; ok: true
claim: Unclear 3,177 and 98.1% coverage; source: 2,999 'unclear' + 178 bodies with no transmitter row = 3,177; ok: true
claim: 35.5% of the brain inhibitory; 91% 'unclear' globally; source: (29,302+22,069+7,891)/166,700 = 0.3555; 0.910 of all transmitter rows; ok: true

Discrepancies

Edge (25,582,938) and contact (124,177,617) counts were not re-counted: that needs the 1 GB edge table, too heavy while the Mac runs science.
The connectome now in data/ (malecns_full.json, 2026-09-16) is a later extraction with a different node policy (164,506 nodes, 25,135,527 edges), so E6's 166,700-node asset is not the file on disk.

Status the catalogued status fits the text

Node count and every transmitter number recompute exactly from the raw MaleCNS tables.

From runs/site_data/exp_checks/E6.json; the log itself is not edited.

Results

EXPERIMENTS.md · line 183

Result.

The full record

EXPERIMENTS.md · lines 174–204

E6 — The whole MaleCNS asset, and what the importer was discarding

Date 2026-09-12 · Question What does the release actually contain, against what the importer kept? · Provenance forager/connectome.py (full_asset_v3)

Method. Node policy "every body with an assigned superclass"; no contact threshold; vectorised aggregation of the 1.05 GB edge table by sorted-index lookup and np.unique/bincount rather than a per-edge Python dict.

Result.

	previous importer	full asset
nodes	163,972	166,700
directed edges	6,143,838	25,582,938
synaptic contacts	88,606,727	124,177,617
extraction time	253 s	11.5 s

Applying the old min_contacts >= 5 rule to this extraction gives 6,242,118 edges, matching Codex's published MCNS count — an independent check that the extraction is correct. The threshold was discarding 76% of all connections.

Transmitters, restricted to the 166,700 retained neurons (98.1% coverage): acetylcholine 103,720 (62.2%), glutamate 29,302 (17.6%), gaba 22,069 (13.2%), histamine 7,891 (4.7%), dopamine/octopamine/serotonin 541 (0.3%), unclear 3,177 (1.9%). 35.5% of the brain is inhibitory. The global table is 91% "unclear" only because it is dominated by unproofread fragments.

Interpretation. Scale was never the obstacle; a Python dict with one entry per edge was. The previous configuration discarded three quarters of the graph and then randomised the signs it had thrown away the evidence for.