Abstract
CI-003 found 46.7% of the trial team's seat-time spent dead. This step finished and measured the instruments that name what a seat was doing when it went down, with no change to play. Over 16 seeds and 819 deaths, two situations account for 63% of deaths in 7.7% of the time: the shell dodge (43%) and the boat's run for the beach, landfall (20%). Afloat, one dodge in six ends in a kill. Because a tank dodges only when a shell is on it, the count cannot say whether the dodge helps. The causal test (the dodge off, and dodge_hold 1, each at 48 seeds) is running and belongs to the next check-in.
Index terms: attrition, reflexes, instrumentation, survival analysis, experimental design.
I. Introduction
A. Intent
To finish the work left in the tree for backlog 28, run its look, bring the brain overview up to date, and commit.
B. The prompts
we are working on the backlog and the roadmap and writing lab reports and researching and implimenting best ai structured brain. look about brain overview in this latest rewrite. finish. commit continue.
commit. push. continue backlog
"Brain overview" was read as LR 20, "The brain now", and BRAIN-MAP.md; "this latest rewrite" as the uncommitted dodge and death-census work. "Finish" was read as build, test, measure and document it.
C. State at the start
Commit a7a6f16 (CI-003, journal 0051), no tag (the Actions pause). Six brain files changed in the tree, not built or measured.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | "finish": work in the tree, untested | Build and test it first; read the diff whole | Committing it as found |
| D2 | make brainmap: two new functions undocumented (backlog 26) | A doc line each | Leaving the count to rise |
| D3 | Backlog 28's proof: a 16-seed census | The null run with CENSUS_ACTS=1; play unchanged, so it also checks the instruments read 0 | A trial run: nothing to try until the look names a part |
| D4 | F: the dodge last held at 43% of deaths | A causal test before any change: -dodge and dodge_hold 1 at 48 seeds | Changing the dodge on the count alone: a reflex that answers fire precedes deaths whether or not it helps |
| D5 | F: 132 deaths a boat sunk by direct fire on landfall | Backlog 30, a weight that turns a boat off a gun's range | Fixing it in this step: one step at a time |
| D6 | "look about brain overview" | LR 20 updated in place: 87 weights, the dodge's weights in the diagram, the 19 October numbers and where it dies | A new overview section: LR 20 is the one the others point to |
| D7 | "commit. push." with the causal runs unfinished | Commit the look as its own step; the runs are the next check-in | Waiting 25 minutes to fold them in |
III. Methods
A. Census
naval.map, 900 s, respawns 3, the fair census (every team in the trial seat on each seed), 16 seeds from 0, the ledger off, CENSUS_ACTS=1. The trial and baseline settings are the defaults (a null run).
B. The parts explored, function by function
| File | Function or field | What it does |
|---|---|---|
fleet.rs | killed_by (new, test helper) | from a tick's events, the weapon whose hit names the tank (pillbox, mine, strike, longshot, direct, indirect), whether it sank, and whether the shooter was the lone RuleBot or no one |
fleet.rs | brain_census: window, died_last, died_window, died_by, died_both | each trial seat's last five codes, one a second while alive; at each TankDestroyed of a trial seat, the four tables |
lib.rs | Brain::dead (new, default nothing), run_bots | called each tick a seat is dead, in place of a think |
lib.rs | Dodge::watch (new), Dodge.watch, metered, tune, began | the dodge under way from its start to its shell's landing; a hit there counted, killed or not, by the dodged shell or not |
lib.rs | Dodge::order_with | the horizon, margin and clearance from tune (the constants when None) |
lib.rs | Dodge::holding | held while driving, and by dodge_hold that share of the flight once arrived |
commander.rs | set_settings, impl Brain::dead, the afloat branch of think_on | the weights into the dodge; the dead seat's watch; Act::Landfall when afloat with no hop under way |
meter.rs | Meter.dodges, dodge_hits, dodge_kills, dodge_shell; dodge, dodge_hit | the counts, ashore and afloat, in the meter's line |
status.rs | Act::Landfall, code landfall | the run for the beach; unique among codes (test) |
weights.rs | dodge_horizon 4, dodge_margin 0.8, dodge_clear 2.5, dodge_hold 0 | the dodge's numbers, after Paul's rule that an either/or is a scalar |
C. Reading the tables
A code's share of deaths over its share of time is how much more likely a seat is to die while holding it. A reflex raises this ratio by being where danger is, so the ratio is a lead for a causal test, not a cause.
IV. Results
A. Deaths (819, 17.1 a match)
| Code held last | Share of deaths | Share of time | Ratio |
|---|---|---|---|
| dodge | 43.2% | 4.5% | 9.6 |
| landfall | 20.1% | 3.2% | 6.3 |
| hop.sailing | 7.9% | 4.7% | 1.7 |
| resupply | 3.9% | 0.9% | 4.6 |
| hop.holding_launch | 3.4% | 1.1% | 3.1 |
By weapon (791 in the twelve commonest causes): indirect 275, direct 253, strike 246, pillbox 17; sunk 399. The commonest pair: direct fire on a boat at landfall, 132.
B. The dodge
Ashore 7,911 dodges: 34.0% hit before the shell landed, 16.6% by the shell dodged, 7.4% killed. Afloat 3,408: 18.2% hit, 7.6% by the shell, 16.6% killed.
C. Null and tests
Trial 8.38, baseline 8.38, 0.00 on every seed. 24.8 us a seat-tick. Brain 69 passed and 8 ignored; client 31; net 14; proto 3; sim 50. No warnings.
V. Discussion
- Two situations, not ten. The hypothesis named two or three, and two make 63% of deaths. The ladder, rotation and the fleet, the object of most steps so far, are under 3% of deaths each.
- The dodge is where danger is. Its ratio of 9.6 is partly that a tank dodges only when a shell is on it. The meter adds that a sixth of ashore dodges are still caught by their own shell, which the clearance (
dodge_clear) and the hold (dodge_hold) speak to. The switch off says whether the dodge saves more than it loses. - Afloat, a hit is a kill (567 of 621). Anything that keeps a boat out of range pays more than anything that makes it hard to hit, which points at the landfall (backlog 30) and the crossings.
Next: read dodge_hold 1 and -dodge against the null (48 seeds each, running); then backlog 30.
Appendix A. Reproducing this report
CENSUS_ACTS=1 CENSUS_SEEDS=16 CENSUS_MAP=$PWD/maps/naval.map CENSUS_RESPAWNS=3 \
CENSUS_SECS=900 CENSUS_LEDGER=0 \
cargo test -q -p treadcore-brain --release -- --ignored brain_census --nocaptureThe causal pair: CENSUS_WEIGHTS a dump of the defaults (CENSUS_DUMP) with dodge_hold = 1, and CENSUS_CAPS=-dodge, each over CENSUS_SEED0 0, 16 and 32 with CENSUS_SEEDS=16.