Abstract
CI-004 found the shell dodge last held at 43% of the trial team's deaths, and that a sixth of ashore dodges are caught by the shell they dodged. This step tested one cause: tanks dropping the dodge on arrival and driving back under the shell. Holding them at the dodge point until the shell lands (dodge_hold 1) costs 1.18 ± 0.55 a seat over 48 seeds and does not change the dodged shell's hit rate (16.7% against 16.6%). The hypothesis is rejected; the weight stays 0.
Index terms: reflexes, causal test, paired census, weights.
I. Introduction
A. Intent
To run the first of the two causal tests CI-004 set, and record it as its own step.
B. The prompt
commit. push. continue backlog
Given after CI-004 was pushed, with the causal runs under way. "Commit. push." was read as recording the finished run now rather than waiting for the second.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | CI-004 D4: a reflex's death share is not a cause | Test the weight that targets the measured failure (the dodged shell's hits) | Tuning all four dodge weights at once: one cause at a time |
| D2 | The prompt, with the hold run done and -dodge not | Commit the hold result alone | Waiting about 20 minutes for both |
| D3 | F: -1.18 ± 0.55, outside two errors; by-the-shell rate unchanged | dodge_hold stays 0; hypothesis rejected | Keeping it for afloat only: the gain afloat is diluted and unmeasured for the trial team alone |
| D4 | F: the hits come during the drive | The next dodge levers are dodge_horizon and dodge_clear, for the tuner | Writing a new dodge before -dodge says whether the dodge pays at all |
III. Methods
CENSUS_DUMP writes the default settings; dodge_hold = 0 changed to 1 is the trial's CENSUS_WEIGHTS. Three runs side by side, CENSUS_SEED0 0, 16 and 32, CENSUS_SEEDS=16, naval.map, 900 s, respawns 3, the fair census, CENSUS_LEDGER=0. The 48 per-seed differences give the mean and its standard error. The meter's dodge counts sum every Commander's dodges in each run.
The part under test is Dodge::holding (lib.rs): with dodge_hold h, a dodge is held while the tank drives to its point, and also while the tick is before began + h * (until - began); at 1, until the shell lands.
IV. Results
| d a seat | Ashore: hit, by the shell, killed | Afloat: hit, by the shell, killed | |
|---|---|---|---|
| Null (CI-004) | 0 | 34.0%, 16.6%, 7.4% | 18.2%, 7.6%, 16.6% |
dodge_hold 1 | -1.18 ± 0.55 | 34.2%, 16.7%, 7.3% | 15.5%, 6.0%, 13.9% |
By range of 16 seeds: -2.02 ± 1.05, -0.38 ± 1.09, -1.14 ± 0.68.
V. Discussion
The meter was built to test exactly this, and it did: the dodged shell's hits did not move, so tanks are not caught after arriving. The cost in score is the errand time the hold gives up. What is left for the dodge is when it starts and where it goes, and first whether it should happen at all.
Next: -dodge at 48 seeds (running); then backlog 30.