Abstract
Backlog 48, the starting numbers for the brain with every part on. LR 55 read every part on at -0.85 +/- 0.59 a seat against LR 54's defaults, and the hypothesis put that cost on one or two parts, by history the posture's lean and team_rotate. A new tool, make parts, turns each of 31 parts (17 capabilities, 14 weights) off from the all-on brain over the fair census at 12 seeds. One part reads past two standard errors, the dodge, and it pays (+5.20 +/- 1.91). The posture (+1.82) and team_rotate (+0.83) read as paying. The plan's halved weights (w_posture 0.25, swap_gap 0.6) read +0.61 +/- 0.60 at 48 seeds, and the table's lowest row halved (landfall_wait 5) -0.09 +/- 0.46. The hypothesis is not supported; the defaults stand. No brain code changed.
Index terms: tuning, ablation, census, every part on.
I. Introduction
A. Intent
To take backlog 48 now that CI-026's feature list is built: find which part, if any, carries the cost of every part on, and set its number.
B. The prompts
continue. also . good luck claude, you are a force of good.
continue
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | The prompt; CI-028 named 48 next | Backlog 48 | — |
| D2 | make switches turns off parts on; every part is on since LR 55 | make parts, its reverse, a script beside it | Reusing switches.sh with a flag: the baseline and trial swap seats and the list differs; two short scripts read plainer |
| D3 | A weight has no -name | Each weight off at the value its doc line names as before it | Halving each: a half-weight measures the number, not the part; the number follows the part |
| D4 | Four cores | JOBS=4, 12 seeds a part, about 1.5 h | 48 seeds a part: about 6 h for 31 |
| D5 | F: no part costs past two errors; the posture and team_rotate read paying | The plan's halved weights at 48 seeds anyway | Dropping them: the plan set the proof before the table, and a 12-seed read is not a refutation |
| D6 | F: landfall_wait the lowest row | Halved (5) at 48 seeds beside it | kite off: a yes/no, and the remaining rows go to 53 together |
| D7 | F: +0.61 +/- 0.60, -0.09 +/- 0.46 | The defaults kept | Taking +0.61: one error; LR 54's rule promotes past two |
| D8 | The table ranks parts below zero (kite, duck, dodge_fire) | Backlog 53, each at 48 seeds | Running them now: the step's proof is done; each is its own question |
III. Methods
Tool. scripts/parts.sh, make parts. A frozen census binary (census_bin); on.weights is caps = baseline; for each part a weights file with caps = baseline, -name or caps = baseline and name = value. The census brain_census with CENSUS_WEIGHTS the trial (every part on) and CENSUS_BASE_WEIGHTS the part off, CENSUS_LEDGER=0; JOBS at once. A table of the trial and baseline mean, the difference a seat and its error, from each log. Env: PARTS, SEEDS (12), SECS (900), MAP (naval.map), RESPAWNS (3), JOBS.
Each part off. JOBS=4 make parts: the 17 capabilities and the 14 weights w_smooth dodge_plan dodge_release dodge_every dodge_fire duck_plan landfall_shy landfall_wait landfall_memory w_front front_memory w_front_post w_front_push w_site_beach, each at 0. naval.map, 900 s, 3 respawns, seeds 0 to 11. build/parts/20261021-085753.
The candidates. build/lr63a/run.sh, build/lr63b/run.sh: a frozen binary each; the fair census, naval.map, 900 s, 3 respawns, seeds 0 to 47 in three jobs of 16, run one after the other; trial (a) w_posture = 0.25, swap_gap = 0.6, (b) landfall_wait = 5; baseline the defaults. Pooled: the mean of the 48 per-seed differences, its error their standard deviation over the square root of 48.
IV. Results
The 31-row table is in section 63. Summary:
| Measure | Result |
|---|---|
| Parts past two errors at 12 seeds | 1 of 31: the dodge, +5.20 +/- 1.91 (pays) |
| Parts reading above zero | 21 of 31 |
| The posture off | +1.82 +/- 0.94 (pays) |
team_rotate off | +0.83 +/- 0.88 (pays) |
| The lowest rows | landfall_wait -2.03 +/- 1.29, kite -1.76 +/- 1.04, duck -1.57 +/- 0.96 |
escort off | 0.00: no escort arose in 12 seeds on naval.map |
(a) w_posture 0.25, swap_gap 0.6, 48 seeds | +0.61 +/- 0.60 (1.0 errors) |
(b) landfall_wait 5, 48 seeds | -0.09 +/- 0.46 (0.2 errors) |
Tests: brain 83 passed and 11 ignored; client 32; net 14; proto 3; sim
- Seven families, each above its floor. 101 weights, 17 bits,
unchanged.
V. Discussion
- The cost is not a part. LR 55's -0.85 was 1.4 errors. Read with a table in which no part costs past two errors and two thirds read as paying, it is most likely the sum of small effects of both signs, or noise.
- The parts the history blamed pay. The posture's lean and
team_rotatecost in earlier, partial brains; with the board, the front and the dodge beside them they read as gains. - Twelve seeds rank; 48 decide. The table is a list of where to spend the night's censuses, which backlog 53 takes.
- A part that never acts.
escortreads exactly 0 on naval.map; its proof needs a map where a builder crosses.
For Paul: none blocking. Backlog 53, the bottom of the table at 48 seeds, is next in goal 4. Commits are local; nothing was pushed.