Abstract
The brain had no weight for spending: shells went at its target tank, longshots every 30 s at structures only, one strike drone at a time. Four weights let it spend, and a cowboy profile at loud values was plain to see. It also won by a margin no change had come near: +11.54 a seat on its census, +11.62 +/- 0.45 read again over 400 seeds, every map above. Taken apart, the longshots carry the fighting and shelling pillboxes is mostly LR 84's reward for it. retreat_shells does nothing at the defaults, because the retreat weights act through w_smooth 0.05. By the rule the spending became the defaults. make families then caught the dodge families at 91-96.5%: a pillbox shot turned the hull into a shell's splash. A pillbox shot now waits under an incoming shell, and the defaults as changed read +11.64 +/- 0.43 against the old.
Index terms: personalities, brain census, ablation, regression, defaults.
I. Introduction
A. Intent
Personalities with code, read by the census: the cowboy first, as Paul named it.
B. The prompts
more census. with the code changes. continue. resume
nice!. looks like its ending. pick up the results.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | "with the code changes" | The cowboy first: four spending weights, at the old behaviour by default | Timid and aggressive first: they needed a cause found (later: the retreat blend) |
| D2 | The machine plays committed trees; nothing committed | A worktree on the machine at 3e55c83 with the patch applied (treadcore-lr86, -lr86b) | Commit to run: Paul has not asked |
| D3 | A weight added must not change play at its default | The same atlantic census to the digit | — |
| D4 | F: +11.54, a style past two errors above | Read again on 400 fresh seeds, and each part alone, rules first | Adopt on 110 seeds |
| D5 | F: re-read +11.62, every map | The spending the defaults; steady.weights the old | — |
| D6 | F: retreat_shells exactly 0 | Left at 4; the blend through w_smooth found and reported | Keep it in the change: it changes nothing |
| D7 | make families: four dodge families under 99% | shell_boxes alone the cause; a pillbox shot waits under an incoming hostile shell | Lower the floors; drop shell_boxes: the fix kept its gain |
| D8 | The match test lost its pings on one seed | The test's teams on steady's spending, as LR 77 pinned mechanism tests | Change the seed: hides the cause |
| D9 | The fix changed the code after the census | The defaults as changed read against steady on 400 fresh seeds, rule first | Adopt on the earlier reading |
III. Methods
The weights in Tactics (from_weights); RuleBot's fight: the pillbox shot after the gun and strikes, the longshot's target, the strike cap. meter::spend on each think's commands. Censuses on the machine, every map, 900 s, 24 jobs: the cowboy (110 from 120000), the re-read (400 from 130000), the parts (110 from 140000), the pair (400 from 150000, BASE=steady.weights). scripts/traits.py; the old score as score less rewards. On the Mac: cargo test -p treadcore-brain --release, make families with SET for each part.
IV. Results
In LR 86: the cowboy +11.54 +/- 0.89; re-read +11.62 +/- 0.45; the parts (shell_boxes +7.19, longshots +5.59, strikes +1.46, retreat_shells 0); the families; the pair +11.64 +/- 0.43, +5.52 without the rewards.
V. Discussion
The brain carried drones and shells it did not use. Spending them is worth more than any weighting of the ladder so far. Part of the measured gain is the reward for pillbox damage that LR 84 added an hour before, but even on the old score the change is four times turtle's. The families earned their place again: a census cannot see a tank shooting itself into a splash on 4% of seeds. The retreat weights acting a twentieth of the way is a finding for every profile that set them.