Lateral inhibition by the strongest file, and a capped age term, 2026-10-18 #

SIMULATED, except one table that is COMPUTED with core's own functions. A review of the coordination math found two things in how ColonyDO.take weighs claims and age (packages/core/src/threshold.ts):

1. Lateral inhibition summed claims over every shared file. An open intent was skipped when other agents' live claims on its own files, added up over all of them, reached half a fresh claim. A claim decays with tau = 5 minutes, so a silent agent's claims on n of an intent's files kept it inhibited for 5·ln(2n) minutes: 3.5 for one file, 18.4 for twenty. Now (FT_INHIBIT_STRONGEST) the strongest single file's claim counts, so inhibition lasts one half-life however many files overlap. Several agents' claims on the same file still add up on that file.
2. The age multiplier had no cap. Stimulus is priority × (1 + ageMinutes / 10) − repulsion, so an intent an hour old had a multiplier of 7 and co-change or import repulsion stopped mattering for it. Now (FT_AGE_CAP) the multiplier stops at AGE_FACTOR_CAP = 4, reached at 30 minutes. The cap was chosen from the sweep below.

Both ship behind flags that are off when absent (packages/worker/src/colony/takemath.ts); off, take() behaves as before. The simulator (packages/bench/src/sim/fire.ts) gained the same two options (inhibitBy, ageCap) and records how long each task waited open before its first claim (waitP50S, waitP99S, waitMaxS).

How these files were made #

node packages/bench/src/sim/inhibition.ts --steps silent --out packages/bench/results/2026-10-18-inhibition
node packages/bench/src/sim/inhibition.ts --steps agecap --variants group+cc5+inh+max,group+cc5+imp1+inh+max --caps inf,2,3,4,6 --out packages/bench/results/2026-10-18-inhibition
node packages/bench/src/sim/inhibition.ts --steps sim,ttz --agents 100,1000 --seeds 1,2,3 --ttz-agents 100,1000 --ttz-seeds 1,2,3,4,5 \
  --variants ft+cc5+inh,group+cc5+inh,group+cc+inh,group+cc+inh+max --ttz-variants ft+cc5+inh,group+cc5+inh,group+cc+inh,group+cc+inh+max \
  --out packages/bench/results/2026-10-18-inhibition/before
node packages/bench/src/sim/inhibition.ts --steps sim,ttz --agents 100,1000 --seeds 1,2,3 --ttz-agents 100,1000 --ttz-seeds 1,2,3,4,5 \
  --variants ft+cc5+inh+fix,group+cc5+inh+fix,group+cc+inh+fix --ttz-variants ft+cc5+inh+fix,group+cc5+inh+fix,group+cc+inh+fix \
  --out packages/bench/results/2026-10-18-inhibition/after
node packages/bench/src/sim/inhibition.ts --steps summary --out <each of the three folders>

Same workload and settings as 2026-10-12-inhibition: hono's history as 100 package copies, co-change table h200, steady state over 20 simulated minutes (5 of warm-up) with a backlog of 1,000, draining 1,500 tasks that are all open at t = 0. Variant names: ft is one commit per lane batch, group the group-commit trunk writer (as deployed); cc is co-change repulsion at weight 2 (the 2026-10-12 runs), cc5 at the deployed FT_COCHANGE_WEIGHT of 5; imp1 adds import repulsion at weight 1 (FT_IMPORT_WEIGHT, off as deployed); inh is lateral inhibition; +max inhibits by the strongest file; +fix is both fixes. The before runs reproduce 2026-10-12 exactly where they overlap (group+cc+inh at 1,000 agents: 175.9 landed a minute, 31.2% rework; draining, seed 3 in 1,563 s and seed 5 in 1,697 s).

Files: silent.json, sim-agecap.json and summary.json here; before/ and after/ each hold sim-steady.json, sim-ttz.json and summary.json.

A silent agent's claims (COMPUTED) #

How long an intent stays inhibited when a silent agent (no heartbeats) holds claims on n of its files, checked every second with core markValue, directClaimOn and inhibited at the deployed threshold of half a fresh claim.

Two silent agents on the same single file hold it 6.93 minutes either way: claims on one file still add up.

Before and after (SIMULATED) #

Mean over seeds (3 in steady state, 5 draining); "slowest wait" is the mean over seeds of the longest any task waited open before its first claim.

• Inhibition by the strongest file changes nothing a live-agent simulation can show, by construction: live agents heartbeat, so every claim in the model is a fresh one (1), and a candidate has a claimed file at 0.5 or more whether claims are summed or not. The +max runs in before/ are identical to the summed ones, run for run. What it changes is the silent case above.
• The age cap at 4 binds only on tasks older than 30 minutes, which the deployed configuration rarely has: it moved only the 100-agent drains of seed 4, where the last task landed 1 s sooner in both group-commit models, and in the per-lane model one task's first claim came 3,240 s after it opened instead of 3,119 s, with the same drain time.

Choosing the cap (SIMULATED) #

Group commit with co-change at the deployed weight 5, alone and with import repulsion at weight 1, both inhibiting by the strongest file; caps of 2, 3, 4 and 6 against none.

Steady state, 1,000 agents, mean of 3 seeds:

Draining 1,500 tasks, 5 seeds (median and slowest of the five drains):

• The cap is about the tail. In steady state no task waits long enough (the slowest wait is under 20 minutes, a multiplier under 3) for 3, 4 or 6 to bind. Draining a fixed set, old intents are what the cap is for.
• With import repulsion on, a cap of 4 shortened the slowest drains: 3,743 → 2,360 s at 1,000 agents and 3,201 → 2,549 s at 100 agents, with the median unchanged and rework a little lower (13.3% → 13.1%; seed 1 at 1,000 agents went from 14.2% to 13.1%). Uncapped, an old intent whose imported files were claimed outgrew its repulsion, was taken anyway, and came back; capped, it waited for those files to land. A cap of 6 never bound there.
• Lower caps cost the anti-starvation purpose. With co-change alone, 2 and 3 slowed the median drain (1,801 → 2,020 and 1,897 s at 1,000 agents; 2,029 → 2,409 and 2,277 s at 100) and lowered no rework: once every old intent has the same multiplier, age stops ordering them. 4 left every co-change-only run as it was uncapped, to within a second.

So 4 is the smallest cap that costs nothing where repulsion is co-change only and keeps import repulsion working on old intents. The model's limits stand as in 2026-10-12: one workload (hono's history), simulated agents with fixed timings, and claims that are always fresh while an agent works.

In this folder #

• README.md
• after/
• before/
• silent.json
• sim-agecap.json
• summary.json