Abstract
Backlog 53. LR 63's 12-seed table read three parts below zero, none past two errors: kite, duck and dodge_fire. Each was turned off from the all-on brain over the fair census at 48 seeds with make parts. The kite off reads -0.94 +/- 0.77, the duck off +0.12 +/- 0.60, the held shot off -0.19 +/- 0.58. None is past two errors, and the defaults stand. The kite keeps a lean at 1.2 errors, but the meter counts no kiting, so the next step counts it (backlog 54). No code changed.
Index terms: tuning, ablation, census, kite, duck, dodge.
I. Introduction
A. Intent
To settle the three lowest rows of LR 63's table at the seed count that can decide them.
B. The prompt
yes. kite , duck, dodge_fire . please try.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | The prompt | kite, duck, dodge_fire off at 48 seeds | landfall_wait: halved, LR 63 read it at -0.09 +/- 0.46 |
| D2 | Four cores, three parts | make parts with SEEDS=48 JOBS=3, one job a part | Three jobs of 16 a part as LR 63a: nine processes on four cores, no faster |
| D3 | F: none past two errors | The defaults kept; no family | Turning the kite off at 1.2 errors: LR 54's rule promotes past two |
| D4 | F: the meter counts no kiting | Backlog 54, the kite counted, before any tuning of its numbers | Tuning kite_margin now: a number for a part not known to act |
III. Methods
PARTS="-kite -duck dodge_fire=0" SEEDS=48 JOBS=3 make parts (scripts/parts.sh, CI-029): a frozen census binary; the trial seat on the defaults (every part on), the baseline with the part off; naval.map, 900 s, 3 respawns, seeds 0 to 47; build/parts/20261021-123214. The difference a seat and its error are the census's own, over the 48 per-seed differences.
IV. Results
| Part off | 48 seeds | Seeds above zero | 12 seeds (LR 63) |
|---|---|---|---|
-kite | -0.94 +/- 0.77 | 22 of 48 | -1.76 +/- 1.04 |
-duck | +0.12 +/- 0.60 | 22 of 48 | -1.57 +/- 0.96 |
dodge_fire=0 | -0.19 +/- 0.58 | 26 of 48 | -1.42 +/- 1.22 |
Tests and families as CI-029 (no code changed).
V. Discussion
- Regression to the mean. The three lowest of 31 rows at 12 seeds were picked for being low; at 48 two of them read near zero.
- The kite's lean is the one left, and a count of it must come before a number for it.
For Paul: none blocking. Backlog 54, the kite counted, is next. Commits are local; nothing was pushed.