Abstract
The shell dodge was the code held last at 43% of deaths (CI-004), and holding it longer cost 1.18 a seat (CI-005). This step turned it off for the trial team over 48 seeds: -2.45 ± 0.67 a seat, outside the noise. The dodge pays. A rough counterfactual from the meter shows it nearly halving the dodged shell's hits ashore and changing nothing afloat. Backlog 28 closes. Backlog 31, a weight for dodging afloat, is the next part.
Index terms: reflexes, ablation, counterfactual, paired census.
I. Introduction
A. Intent
To finish backlog 28 with the ablation CI-004 set.
B. The prompt
commit. push. continue backlog
Read as: record the dodge-off run when it lands, commit and push it, and set the next backlog item.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | F: -2.45 ± 0.67, 3.7 errors | The dodge stays on; backlog 28 done | Reading the 43% death share as a cost: the ablation says the opposite |
| D2 | F: the meter still counted near-null dodges with -dodge | Read the code: order_with meters before dodges drops the order; use it as a would-be count | Calling the run broken: the trial team's moves were dropped, as the score shows |
| D3 | The would-be counts are mixed with the baseline teams' | Estimate undodged rates as 3 × run − 2 × null, labelled as a direction | Reporting the mixed rates as the trial team's |
| D4 | F: afloat, undodged ≈ dodged | Backlog 31: dodge_afloat, a weight (1, today's), after the either/or rule | Turning the boat's dodge off outright: a scalar lets the tuner find the value |
| D5 | D3's estimate | Backlog 31 starts with the meter by team, would-be apart from taken | Building dodge_afloat on an estimate |
III. Methods
CENSUS_CAPS=-dodge (the trial team's dodges false; Commander:: set_settings), the defaults for the baseline teams. CENSUS_SEED0 0, 16, 32 with CENSUS_SEEDS=16, side by side; naval.map, 900 s, respawns 3, the fair census, the ledger off.
The part: Commander::think_on calls self.dodge.order_with(..) (lib.rs), which starts the watch and meters, and then filters the order by self.dodges; holding is likewise gated by dodges. With the switch off, the meter therefore records each dodge the trial team would have made and its outcome undodged.
Estimate: with the trial team's share of shells taken as one third, run rate r = r_u / 3 + 2 r_d / 3, so r_u = 3r − 2r_d, with r_d the null run's dodged rate (CI-004).
IV. Results
| d a seat | |
|---|---|
dodge_hold 1 (CI-005) | -1.18 ± 0.55 |
-dodge | -2.45 ± 0.67 |
By range of 16 seeds: -4.11 ± 1.10, -1.39 ± 1.14, -1.84 ± 1.19.
| Dodged (null) | Undodged, estimated | |
|---|---|---|
| Ashore, hit by the shell | 16.6% | ≈30% |
| Ashore, killed | 7.4% | ≈11% |
| Afloat, hit by the shell | 7.6% | ≈8% |
| Afloat, killed | 16.6% | ≈14.5% |
V. Discussion
- A death census overstates a reflex. The code most deaths passed through is the part worth most. Before reading any reflex's share of deaths, the ablation has to be run.
- Ashore and afloat are two parts. The same dodge saves a tank and costs a boat its crossing for nothing. That is an either/or, so it becomes a weight.
- Instrument by team. The meter's per-world counts forced an estimate; per-team and taken-or-not counts make the next reading exact.
Next: backlog 31, the meter first; then backlog 30.