Abstract
Every census since LR 11 played naval.map alone (CI-001 A8). The census now plays a list of maps, MAP=all being all eleven, and reads a difference on each beside the pooled one. Its first reading puts the dodge, on against off, over 176 seeds: +5.94 +/- 0.38 a seat, 15.6 errors (read first part-way, 123 seeds, +6.00 +/- 0.48). Ten maps of eleven are past two errors for the dodge, and none against it. The dodge holds by the rule set before the run. Naval.map is mid-table, so the one map did not overstate the part.
Index terms: brain census, maps, paired difference, standard error, dodge.
I. Introduction
A. Intent
Backlog 25: does a part's difference on naval.map hold on other maps? The dodge, the largest part (LR 29, LR 63), is the first one asked.
B. The prompts
The census was built and launched earlier in the session. Paul then stopped some runs and asked:
i just cancled some drive runs mid process. do you have record o the last run. does it give any insite?
The census in build/census/20261021-193635 had not been stopped; it was read part-way (D2). Then:
ok. so update. commit. push. i like this.
treadcore build census keeps running. whats its intention.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | CI-001 A8: one map | CENSUS_MAP a list or gen; seed n on map n mod k, so jobs keep a seed's map | A map per job: a job's maps would differ with JOBS |
| D2 | Paul asked for insight before the census ended | Read part-way from the per-match team lines; the census's own pooled.txt stays the record | Wait an hour, or stop it: he asked for it now |
| D3 | "update. commit. push." | Write LR 71 on 123 seeds, marked part-way, and push | Hold the write-up for 176 seeds |
| D4 | F: MAPS.md's paragraph on how matches end came from one seed a map and is wrong on naval and the fortress isles | Corrected from the 372 matches | — |
| D5 | F: a comment named scripts/maps.sh | Named scripts/pool.sh, the script that pools | — |
| D6 | A cached .pyc in the tree | Removed; __pycache__/ ignored | — |
III. Methods
The census. scripts/census.sh: the fair census (brain_census), trial weights caps = baseline (every part on), baseline caps = baseline, -dodge; MAP=all, which is the five prior-art maps, the two world maps, industry, island, naval and gen. Seeds 1000 to 1175 (16 a map), 900 s, 3 respawns, four jobs of 44 seeds from a binary copied when the run began; ledger off.
The part-way reading. For each seed, all three of its matches played (each team in the trial seat once): the trial team's mean score a seat, averaged over the three matches, less the baseline teams' mean. Its mean and standard error over seeds, overall and by map; past two errors is outside the noise. The census weights by seat (teams of 6, 5 and 5). This reading weights by team line, so it may differ from pooled.txt in the second decimal. The meter's dodge counts were summed over every match played.
IV. Results
| Dodge on less off, a seat | |
|---|---|
| Every map, 176 seeds | +5.94 +/- 0.38 |
| Highest | island +10.86 +/- 1.48; generated +9.01 +/- 1.19; industry +8.41 +/- 0.95 |
| naval (16) | +4.64 +/- 1.37 |
| Lowest | atlantic +2.99 +/- 0.53; fortressisles2 +2.29 +/- 1.15 (inside) |
Ashore, the dodged shell hits 2.4% of the time with the dodge and 23.0% without. Afloat it hits 0.5% against 4.3%, but deaths after the shell fall less (9.0% against 12.2%). The full table is in LR 71.
V. Discussion
- The dodge holds everywhere. It is the one part so far past two errors on a single map. On eleven maps it is past fifteen errors.
- Where it pays least. It pays least on the island-chain maps, where more of the fight is afloat. This matches LR 29: afloat, the dodge changes little (backlog 31).
- The method. Ten maps cost no more than one at the same seeds. Promotions can now ask for more than one map.
Update, 22 October: the census ended; LR 71 and this report carry its pooled.txt (176 seeds), which replaces the part-way reading.