lab reports · orreryAbstractIIIIIIIVVVIVII

treadcore · check-in report CI-089

The Duelist's Gate a Dial, a Clone Judged Beside the Rules, and crucible's Partners in the Game's Styles

Check-in report CI-089. Paul Richeson, treadcore.org; Claude Code (Claude Opus 5.5) as the pair. 4 November 2026. Companion records: LAB-REPORT-BRAIN.md section 140; CHANGELOG; ROADMAP backlog 82; journal 0116; crucible cab1adf; orrery 7b3bf3c.

Source: documentation/checkins/CI-089-2026-11-04-the-gate-a-dial-the-clone-beside-the-rules.md, built by orrery's tools/report/collect.py.

Abstract

The Gladiator read 2.5 a seat under the game's defaults, and the clone stage's first round could not tell a clone from the Scholar as a whole team of nets. The Duelist's gate is now a dial the census can read (duel_gate); the clone is judged on one seat beside the rules (SCHOLAR=file, bot_mix = one); and crucible's world islands spar against the game's styles. The gate swept at 0, 0.3 and 0.6 leaves the default at 0: 0.6 leans better by +1.61 ± 1.50, inside two errors.

I. Intent

orrery asked what should close the Gladiator's gap and how a clone should be judged. Paul chose, word for word: "Islands meet the game's styles", "The gate as a dial", and "One seat beside the rules".

II. Decision log

DecisionCause
duel_gate, default 0, range −1 to 1CLAUDE.md, "either/or is a scalar"; 0 is the gate as trained; the confidence is a tanh
The gate read from the rules' settingsevery dial of a seat lives there; the Duelist wraps the rules
The gate for both styles 7 and 17one bridge, one rule
SCHOLAR=file for slot 6a census of one seat beside the rules needs the style, not CENSUS_NET's whole team
The gate's default not movedthe rule before the run: no gate past two errors
crucible's styles only under WORLD, style 0 the scriptthe yardstick and the hybrid's answer must stay what they were measured on

III. Method

weights.rs DuelGate; RANGES; styles.rs Duelist::gate, scholar_file, a test that a confidence of 0.54 passes gates 0 and 0.5 and not 0.6. 164 brain tests. Three censuses on this Mac, every map (the BSD paste shim), 48 seeds, 900 s, respawns 3, JOBS=8, the trial bot_style = 17 with duel_gate 0, 0.3, 0.6; paired seed by seed.

IV. Results

GateRead, a seat against the defaultsLess gate 0, paired
0−2.58 ± 1.27
0.3−2.50 ± 1.16+0.08 ± 1.37
0.6−0.97 ± 1.52+1.61 ± 1.50

V. Discussion

A champion kept to the fights it is surest of costs its team less: at 0.6 the Gladiator reads inside the noise of the rules. One error is not two, and 48 seeds are a tenth of a held-out read; the fleet's 400 will say whether 0.6 is the line.

VI. What is not built

The gate's read on the fleet; the clone's first one-seat judgement (it waits for this commit); champions trained against the styles, which the bridge reads as they come.

VII. Paul's decisions

Whether to commit and push this; whether the fleet reads the gate at 0.6 on 400 seeds.