Abstract
Two items from CI-020's conversation. First, the census scripts now run a frozen copy of the test binary, so brain code can change while a census plays. This was proved by breaking the source mid-run, and the run finished with an identical reading. Second, on Paul's choice, the capability smooth is now a weight, w_smooth, with an insignificant default of 0.05. Its numeric parts blend by the weight; its yes-or-no parts act above 0.5. The old play against the new default reads +0.47 ± 0.60 a seat over 48 seeds.
Index terms: tooling, reproducibility, weights, either-or rule.
I. Introduction
A. Intent
Backlog 41 and 40, set at CI-020's close.
B. The prompts
is there an improved and faster way for you to test ?
ok. maybe not remove it. give it an insignificant coefficient
commit. push. continue backlog
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | Each census rebuilt the brain | census_bin: build once, copy the binary beside the logs | A second target directory: the copy also records which binary made a result |
| D2 | No new dependencies | Parse cargo's JSON with python3 | jq |
| D3 | Proof of D1 | Break the source mid-run; compare with a known reading | Trusting it |
| D4 | The proof's wait loop matched itself | Stopped; the source restored from a copy, the tree checked clean | |
| D5 | Paul: "an insignificant coefficient" | w_smooth 0.05 | 0: indistinguishable from removal |
| D6 | The stance's smoothing at small shares is another rule | Thresholded with the counts | Blending it: a small share would change play more, not less |
| D7 | Old files and make switches name smooth | Read as the weight at 1 or 0 | Breaking them |
| D8 | Proof of D5 | The old play as trial against the default, 48 seeds |
III. Methods
census_bin DIR (scripts/common.sh). The smooth change in weights.rs (SmoothShare, smooth, smooth_alpha, smooth_on, the parser), commander.rs (set_settings, go, needs_home, auto) and fleet.rs (the census's capability reader). Census: w_smooth = 0 as the trial weights, the defaults for the baselines, 48 seeds in three ranges, naval.map, 900 s, respawns 3, the fair census, from build/lr46/census-bin.
IV. Results
Frozen binary: the broken-tree run read d −0.4567 ± 0.4054, as LR 40's first plumbing run. w_smooth 0 against 0.05: +0.47 ± 0.60. Tests: brain 74 passed and 8 ignored; client 31; net 14; proto 3; sim 50; no warnings.
V. Discussion
Paul's rule turned a removal into a dial: the six behaviours stay, reachable by one number, at a setting that does not show in play. The frozen binary let this step's documentation be written while its census ran, which is what Paul's question was after.
Next: backlog 39.