The Posture as Scores

Abstract

CI-001 found that the team board's posture, which steers the Auto ladder under the posture capability, was a first-match ladder on thresholds, where the design (LR 9) called for scores. This step replaced it with five scores, each from 0 to 1 and each with a weight. A new scalar, posture_scored, blends them with the old ladder, so the old posture stays exactly one setting away. A unit test shows each posture winning on its own case. In three of the five cases the old ladder would have said Fortify. Measured over the same 48 seeds as CI-001, the scored posture read -1.04 ± 0.58 a seat, against -0.81 ± 0.65 for the ladder's: -0.23 ± 0.74 apart, no difference. Together the two read -0.93 ± 0.49. That moves the suspicion from the posture to the lean that overrides the ladder's order. posture stays off.

Index terms: utility AI, team posture, influence map, paired comparison, standard error.

I. Introduction

A. Intent

To finish backlog 21, the next item CI-001 set out: the posture scored as LR 9 designed it, measured against the ladder it replaces on the same seeds.

B. The prompt

commit and push, then start backlog 21

C. State at the start

Commit 45641fc (CI-001, journal 0049), pushed to main, with no tag, so no workflow ran. posture off and measured at -0.81 ± 0.65 with the ladder's posture. The standing rules of CI-001 I.C hold.

II. Decision Log

III. Methods

A. Apparatus and census

As in CI-001 III.A to III.C: naval.map, 900 s, respawns 3, 16 seats, the fair census, and the paired difference over seeds with its standard error. Three ranges of 16 seeds (CENSUS_SEED0 0, 16, 32) ran side by side. The trial was board, posture with w_posture 0.5 and posture_scored 1; the baseline was board.

B. The parts touched, function by function

C. The scores

With T the hostile worth within the base's 16-tile alarm (a full tank 1), m the members' mean material share, and o the ratio of allied to hostile tank worth in sight (1 if no hostile is seen):

s_D = \min(1, T / 0.5), \quad
s_E = \begin{cases}1 & \text{starved}\\ 1 - m & \text{else}\end{cases}, \quad
s_F = \text{castle\_left}, \quad
s_C = \text{slots\_left}, \quad
s_X = \operatorname{clamp}\!\left(\frac{o - 1}{\text{press\_odds} - 1}, 0, 1\right)

\text{value}(p) = (1 - k)\,[p = \text{ladder}] + k\, w_p\, s_p

The best value wins, ties in the order Defend, Economy, Expand, Connect, Fortify; Fortify if every value is 0.

D. The proof

each_posture_scores_its_own_case: sixteen seats in two teams, ours by the base and theirs 80 tiles off, with full vision (advantage). Five cases are built from the same world, each varying one input, and a sixth puts a hostile tank at the gate. Each case must pick its posture at posture_scored 1 and the ladder's pick at 0. The test prints each case's scores (D6).

IV. Results

A. The test

B. The census (48 seeds)

The seed spread was 4.0 a seat. The two runs' per-seed differences correlate at 0.28: a seed's start explains little of either.

C. Tests

make test, every crate: brain 68 passed and 8 ignored (new: each_posture_scores_its_own_case); client 31; net 14; proto 3; sim

1. No warnings.

V. Discussion

1. The posture was not the problem. Two different ways of working it out give the same result within the noise. Finding A3 of CI-001 was a real design fault, and fixing it changed which postures are read (Economy 16% to 37%, Connect 19% to 6%), but not the score.
2. The lean is the suspect. Both forms override the ladder's own order, and both are about a point behind. The ladder's order (DefendBase, Fortify, Turrets, Hunt, Expand, Guard) is the brain of 17 October, tuned by play; a lean that lifts Expand over Fortify whenever material is low may simply be worse play. Backlog 27 measures which rungs gain.
3. Ranges read their own luck again. The last 16 seeds alone were outside two errors (-2.07 ± 0.99); the 48 are not. Only the whole sample counts, and only on seeds chosen before the run.
4. On the architecture. The step added a Home read-out because the board cannot see the plan. That is finding A5 of CI-001 again: the brain's state is spread across the commander, the plan and the board, joined by whoever is first due. A team-level owner of the plan (the fleet commander, crew.rs) would be the natural place for these inputs. That is a larger change, for when the lean pays.

Next: backlog 27, which rungs the lean moves; then 23, the honest tuner.

Appendix A. Reproducing this report

CENSUS_DUMP=scored.toml CENSUS_SEEDS=0 CENSUS_CAPS=board,posture \
  cargo test -q -p treadcore-brain --release -- --ignored brain_census --nocapture
for k in 0 16 32; do CENSUS_SEED0=$k CENSUS_SEEDS=16 CENSUS_WEIGHTS=scored.toml \
  CENSUS_BASE_CAPS=board CENSUS_MAP=$PWD/maps/naval.map CENSUS_RESPAWNS=3 CENSUS_LEDGER=0 \
  cargo test -q -p treadcore-brain --release -- --ignored brain_census --nocapture > scored-$k.log 2>&1 & done; wait
cargo test -q -p treadcore-brain --release -- each_posture --nocapture