lab reports · orreryAbstractIIIIIIIVVVIVII

treadcore · check-in report CI-090

The Boundary as a Test: Gauntlets G18 to G22

Check-in report CI-090. Paul Richeson, treadcore.org; Claude Code (Claude Opus 5.5) as the pair. 4 November 2026. Companion records: LAB-REPORT-BRAIN.md section 141; CHANGELOG; ROADMAP backlog 83; journal 0117; scenarios/gauntlet/g18 to g22.

Source: documentation/checkins/CI-090-2026-11-04-the-boundary-as-a-test.md, built by orrery's tools/report/collect.py.

Abstract

Five single-decision tests mark the boundary of smart play around pillboxes and attackers: shoot the hostile pillbox and spare the neutral one; take two pillboxes one at a time from outside their rings; answer the tank that shot you before a nearer one that did not; kite an attacker when alone; stand and fight it with friends near. No brain meets them on 50 seeds, the rules included (24, 0, 48, 34, 56%).

I. Intent

Paul, word for word:

we need to setup a goal where it would be smart to attack some pillboxes while avoiding others. like enemy preference , last attacked enemy we would kite and dance, unless we have backup with fleet allies nearby. we should solve it by creating a coded set of tests that should result into that being the boudray of your question.

II. Decision log

DecisionCause
Gauntlet files, not a new harnessthe gauntlet is the single-decision test already, read for every brain with a floor
Hostile: shoot; neutral: sparethe sim: a live hostile pillbox must be shot dead before it is taken, a live neutral one is taken whole and fires on whoever enters its ring
Every pillbox test passable from outside the ringsa gun of ten against a ring of eight; a test the game cannot pass is not a boundary
G20's attacker fires firstthe tank's gun reloads its first two seconds
G21 and G22 as a pair"kite and dance, unless we have backup": the same fight alone and with friends
G22's friends hold their groundescorting friends made staying with them their choice, not the tank's (an untrained net passed 44%)
G22 asks the tank to outscore the attackertwo hits were most of a kill friends share (the rules 14%)
Floors above every brainthe floor is the boundary; a brain earns it

III. Method

scenario.rs: pillbox_struck (PillboxDamaged on a placed pillbox), first_hit (the first tank a seat hit), pillboxes_standing. Five files. The gauntlet on this Mac, 50 seeds from 9,500,000, the rules, the untrained net, the six milestones, bc2 and logistics (GAUNTLET_NETS); each file played by the rules on 20 seeds for the failures in words (make scenario). 164 brain tests.

IV. Results

the rulesuntrainedmap7map5bc2logistics
G18 (floor 0.8)24%0%0%0%0%0%
G19 (0.7)0%0%6%0%0%0%
G20 (0.7)48%0%8%4%0%2%
G21 (0.6)34%2%0%26%0%6%
G22 (0.7)56%30%20%40%16%8%

The rules' failures, in the harness's words: the neutral pillbox struck for 55 to 65 and 45% of the armour gone (G18); "1 placed pillboxes stand at the end" and 55% armour (G19); "seat 0 hit seat 1 first, seat 2 expected" (G20); dead, or hit as often as hitting (G21).

V. Discussion

The rules play these decisions about half as well as they should, and the nets worse: the boundary is real and nobody is at it. G22 cannot tell a passive tank beside its friends from an active one as well as the others tell theirs apart (an untrained net 30%), and is read with G21: a brain that passes both has learned when to kite and when to stand.

VI. What is not built

The styles (Duelist, Gladiator) in the gauntlet; the rules taught the five; crucible's arena given a hostile and a neutral pillbox.

VII. Paul's decisions

Whether to commit and push this; which of the next three comes first.