Abstract
CI-013 found +2.31 ± 0.58 a seat with three landfall weights on together: landfall_memory 30, landfall_shy 1, landfall_wait 10. This step took each back to 0 in turn over 48 seeds. Without the memory, -1.08 ± 0.53. Without the stand-off, +0.69 ± 0.64. Without the run's price, +1.50 ± 0.67. The memory is the base: the other two cost a point without it and add about 2.4 a seat between them with it. All three together is the best setting measured. The defaults are Paul's.
Index terms: ablation, complementary features, memory, weights.
I. Introduction
A. Intent
Backlog 37: which of the three carries the gain, before Paul chooses defaults.
B. The prompts
commit. push. continue backlog
Given twice while the runs went on. The first finished run (memory off) was committed alone (journal 0062); this report covers all three.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | CI-013 D6 | One weight back at 0 in each of three runs, 48 seeds, the same seeds | Each alone from the null: shy and wait alone are LR 31 and 33 already |
| D2 | A prompt mid-ablation | Commit the finished run alone; this report for the whole | Waiting 20 minutes |
| D3 | F: memory off -1.08; the others positive | The memory is the base | |
| D4 | F: all three best; the values hand-picked | Recommend all three on at these values; the tuner later | Choosing the defaults: Paul's |
III. Methods
From CI-013's weights file (memory 30, shy 1, wait 10), three files each with one of the three set to 0. Each run: CENSUS_WEIGHTS, 48 seeds as three ranges of 16 side by side, naval.map, 900 s, respawns 3, the fair census, CENSUS_ACTS. The runs went one after another. Each reads against the null (the defaults for every team) seed by seed.
IV. Results
| Trial | d a seat | Landfall deaths, seeds 0–15 | 48 seeds | Seen near there earlier, 0–15 |
|---|---|---|---|---|
| Null | 0 | 165 | 76 | |
| All three | +2.31 ± 0.58 | 82 | 281 | 16 |
| Memory off | -1.08 ± 0.53 | 164 | 542 | 72 |
| Shy off | +1.50 ± 0.67 | 103 | 341 | 16 |
| Wait off | +0.69 ± 0.64 | 113 | 369 | 25 |
V. Discussion
- Complementary parts. The memory knows where a gun is; the run's price keeps the boat's course off it; the stand-off keeps the boat off a beach it covers. Each needs the memory, and each adds to it.
- Differences between trials are rough. Each run is paired with the null, not with the others, so "the stand-off adds 1.6" carries the errors of two runs (about 0.9).
- The values are hand-picked. 30 s, 1 and 10 s were the first tried. The tuner (backlog 23) can search them once they are on.
Next: Paul's defaults for the three.