lab reports · orreryAbstractIIIIIIIVV

treadcore · check-in report CI-065

The School: Both Brains Training at Once, and How More Machines Would Help

Check-in report CI-065. Paul Richeson, treadcore.org; Claude Code (Claude Opus 5.5) as the pair. 27 October 2026. Companion records: LAB-REPORT-BRAIN.md section 103; CENSUS.md section 5a; build/school/NOTES.txt on the Mac.

Source: documentation/checkins/CI-065-2026-10-27-the-school.md, built by orrery's tools/report/collect.py.

Abstract

The learned tank now trains continuously beside the scripted brain: make school runs its curriculum until stopped on the Mac, as make converge runs the census on the census machine. A round keeps a net only on a real gain with nothing broken, then gauntlets it. Two training maps (low armour with full shells; an unguarded depot) answer the shortcuts the gauntlet found. Each read first at 100% for the scripted brain, 0% for the untrained net and 11% for the map 7 net. More census machines would scale the run by whole censuses from a shared queue; that is planned in CENSUS.md section 5a, to be built when a machine arrives.

I. Introduction

A. Intent

Train both brains at once, and say how more CPUs could be used.

B. The prompts

we will continue to test. are we training both brainz and net ai.

yes. build it and start it on the mac. please. also if I get more CPU. is there a way to have multiple census machines.

II. Decision Log

#CauseDecisionNot taken, and why
D1Only the scripted brain trained continuouslymake school: the curriculum until stopped, on the MacTraining on the census machine: the run fills it
D2LR 102's shortcuts (G5, G6)Maps 9 and 10, farther than their gauntlet tests, so the gauntlet stays a testTrain on the gauntlet's files: it would stop measuring
D3A loop needs its own ruleKept only on a gain of two errors with no passed course under its floor; the rule in NOTES before it playedKeep the newest net each round
D4LR 100: 114 of 117 phase 1 "failures" were killsPhase 1's enemy ground counts a kill as reached, written before the school's first read—
D5"multiple census machines"Whole censuses from a shared queue, the ledger merged; plannedJobs of one census over machines: breaks rule 1
D6The learned tank on several machinesIslands: a school each, the best carried nightlyOne generation over a network: it waits on the slowest

III. Methods

make school from nets/map7.bin, 300 generations a round, held-out reads as in LR 99 to 102, and the gauntlet at 200 seeds a test. Baselines by make train GROUND=... BASELINE=1 CHECK=....

IV. Results

Started; its rounds are read at the next morning read (build/school/rounds.tsv).

V. Discussion

With both loops running, the morning read covers two things: the run's candidates for the scripted brain, and the school's kept nets with their gauntlets.