Co-change repulsion: files that change together wire togetherSIMULATEDMEASURED

On hono's history FireTrident redid more work than locks. Part of it: agents took intents whose files are coupled to files other agents were changing, and read-set validation sent them back. TrunkDO now mines the trunk's history at import into a co-change map (Hebbian: a file's weight to a partner estimates P(partner changes | file changes) and decays only as that file changes again; STDP: an author's follow-up commit, or a follow-up intent, potentiates earlier-file → later-file edges). It learns every landing after that. ColonyDO.take adds 2 × coupling × claim for each claimed neighbor of a candidate's files, so a claim repels work on the files that change with it.

Does the map predict co-change? MEASURED on hono's real history (train on 200 commits, predict the next 100, sliding)

predictor	precision@1	precision@3	recall@8	files with neighbors
co-change map (as deployed)	0.53	0.35	0.29	63%
most changed files	0.13	0.11	0.15	-

Direction: ranking the files of the author's previous commit (the upstream change), recall@8 is 0.032 with STDP and 0.026 without. In hono the follow-up signal is weak (squash-merged PRs from many authors), so STDP mostly orders downstream work after upstream work rather than adding coupling.

hono workload, 20 simulated minutes, mean of 3 seeds (SIMULATED)

agents	model	landed/min	rework %	conflicts/1k	claim to land p50 s
100	Baseline: locks + one merge queue	56.2	5.7	72	76
100	FireTrident as built	62.1	12.3	75	64
100	FireTrident + co-change	63.4	10.0	63	64
100	FireTrident + group commit	62.0	12.7	77	64
100	FireTrident + group commit + co-change	64.5	10.1	65	63
1,000	Baseline: locks + one merge queue	51.9	21.7	318	563
1,000	FireTrident as built	85.6	42.5	397	355
1,000	FireTrident + co-change	85.9	38.7	354	325
1,000	FireTrident + group commit	205	45.3	319	72
1,000	FireTrident + group commit + co-change	192	40.4	279	69
10,000	Baseline: locks + one merge queue	51.9	21.7	318	563
10,000	FireTrident as built	85.8	41.4	386	355
10,000	FireTrident + co-change	85.8	38.6	365	326
10,000	FireTrident + group commit	206	46.9	334	70
10,000	FireTrident + group commit + co-change	192	40.3	279	69

Co-change weight: rework % @ landed/min, mean of 3 seeds (SIMULATED; 0 = off, 2 = deployed)

agents	model	0	1	2	3	10
1,000	as built	42.5% @ 86	39.8% @ 86	38.7% @ 86	37.8% @ 86	35.9% @ 86
1,000	group commit	45.3% @ 205	43.2% @ 197	40.4% @ 192	39.8% @ 186	36.4% @ 154
10,000	as built	41.4% @ 86	40.8% @ 86	38.6% @ 86	37.4% @ 86	36.4% @ 86
10,000	group commit	46.9% @ 206	43.0% @ 203	40.3% @ 192	39.1% @ 187	36.8% @ 152

Draining 1500 tasks to zero, seconds, mean of 10 seeds (median for the last task) (SIMULATED)

agents	model	50% landed	90%	95%	all (median)	rework %
100	Baseline: locks + one merge queue	813	1,455	1,536	1,998	5.2
100	FireTrident as built	702	1,273	1,345	2,376	12.4
100	FireTrident + co-change	694	1,258	1,328	2,360	11.2
100	FireTrident + group commit	693	1,258	1,330	2,477	12.4
100	FireTrident + group commit + co-change	684	1,244	1,312	2,303	11.1
1,000	Baseline: locks + one merge queue	847	1,557	1,644	2,277	15.9
1,000	FireTrident as built	533	953	1,075	2,323	39.2
1,000	FireTrident + co-change	531	952	1,074	2,261	35.8
1,000	FireTrident + group commit	116	411	565	1,703	37.1
1,000	FireTrident + group commit + co-change	131	429	591	1,831	33.4

Live steering on a preview deployment (MEASURED: real import of hono, real mining, real takes)

Each round a fresh agent claims src/jsx/components.ts (priority 5), then a fresh agent chooses between Y (priority 2.5) and src/utils/body.ts (priority 2). In the coupled colony Y co-changes with the claimed file in hono's last 200 commits; in the control colony it never did.

colony	Y	took Y	took Z	other	take p50 ms
coupled	`src/jsx/streaming.ts`	2/20	14/20	4	319
control	`src/router/reg-exp-router/node.ts`	16/20	3/20	1	292

Two-sided Fisher exact test on took Y / took Z, coupled vs control: p = 3.0e-5. "Other" rounds are B declining both (the response-threshold rule) or A not taking X first.

Rework falls, throughput holds on per-lane landing. At 1,000 agents rework drops from 42.5% to 38.7% at the same 85.9 landings/min; at 100 agents from 12.3% to 10.0% with slightly more landings. Locks stay at 21.7%.
With group commit it is a trade. Rework 45.3% → 40.4%, landings 205 → 192/min at 1,000 agents, and draining a fixed backlog slows a little: repelled intents sit in the 200-oldest candidate window until their neighbors land. FT_COCHANGE_WEIGHT tunes it (0 turns it off).
What is left of the gap to locks (SIMULATED analysis, not built): direct overlap is only soft repulsion, so the age term and reinforced thresholds let agents take claimed files anyway. Treating a direct overlap as exclusion takes rework to 33.9%, and adding co-change to 31.2% (weight 2) or 26.5% (weight 10), at 86 landings/min, against locks' 21.7% at 52.
Method. The simulator's take adds the same core neighborRepulsion ColonyDO uses, with neighbors mined from hono's last 200 first-parent commits (TrunkDO's import depth), top 8 per file, coupling ≥ 0.1. Tasks are drawn from the co-change graph of hono's whole history, while the colony's map knows only the last 200 commits, as a fresh import does. The simulated colony does not learn from its own landings (the deployed TrunkDO does).