Abstract
The session first committed CI-026's finished step, after checking its tests (brain 81, client 32). It then took backlog 51, the rest of LR 58's hypothesis: a guard should meet an attack before it reaches the base. Two weights lean the Guard rung toward the team board's front. w_front_push (0.75) moves the rotation's push point that share of the way toward the front sector nearest it. w_front_post (0.5) moves the guard's post toward the nearest front sector, 6 tiles at most. A new family, a raid on a guarded base from any side, meets the raider beyond the base's ring in 85.5% +/- 1.8% of 400 seeds, against 50.5% +/- 2.5% with the parts at 0. All of that gain comes from the push point. The family shows the first form of a raid, one that the guard's drones and a teammate could see, being met 20 tiles out whatever the post. The guard census reads -0.27 +/- 0.64 a seat over 48 seeds, inside the noise, with the push point acting in 79% of a rotating guard's seconds.
Index terms: team board, influence front, rotation, guard post, scenario families, census guard.
I. Introduction
A. Intent
To pick up where the last session stopped, and go on through the backlog in the order CI-026 named: 51, then 20, then 48.
B. The prompt
resume.
The working tree held CI-026's step whole and uncommitted (journal 0083 said "this entry's commit follows"). It was committed locally as c61ea8a after its tests were run again, and not pushed.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | "resume."; CI-026 finished but not committed | Its tests run again, then committed locally, not pushed | Pushing: AGENTS.md, commit or push only when asked; a push is outward-facing |
| D2 | CI-026, next steps: 51 | Backlog 51 | 20 first: CI-026 put 51 ahead |
| D3 | Either/or is a scalar (Paul, 18 October); the post and the push point move by different amounts | Two weights, w_front_post and w_front_push, rather than the one the backlog named | One weight: the tuner could not tell them apart |
| D4 | The guard must still cover the base | The post moves at most BASE_ALARM - BASE_R, 6 tiles | An unbounded lean: the base's ring would leave the leash |
| D5 | The rotation's defence end is where the tank refills | The defence point stays the base's post; only the guard's stand moves | Moving Commander::post itself: it moves the refill point, the window's points and the pillbox carry home |
| D6 | F: a brainless raider was not a raid (the RuleBot went for a far depot) | A tank headed at the base (toward, new) | A scripted Commander raider: more parts in play than the one under test |
| D7 | F: raids off the map's edge | The base at the map's centre ("centre" without a number) | Clamping positions: changes the geometry silently |
| D8 | F: the parts on and at 0 gave the same counts at every bar; the raid met 20 tiles out by shells and drones | The teammate's sight and the drones taken away after half a second; the raid from 40 tiles | Reporting the first family: it measured the drones, not the post |
| D9 | F: the push point does all of it; 0.75 is within two errors of 1 | w_front_push 0.75 | 1: the push point on the front itself, farthest from the base |
| D10 | F: the post did nothing in the family | Kept on at 0.5, its use metered | Off: Paul's rule of 20 October, features on, tuning after (48) |
| D11 | LR 58: count before you census | A meter of seconds leaned, read in the first matches (404 to 537 a match) before the census was let run | The census blind |
III. Methods
Code. weights.rs: FrontPost (w_front_post), FrontPush (w_front_push). commander.rs: front_points (the front's sector centres with Caps::BOARD, now also used by expand); guard_post (the post moved min(w * d, FRONT_POST_REACH) toward the nearest front centre, back to the post if the world refused that tile); defend_base uses it outside an alarm; think_map hands the rotation front and front_w before fit. rotation.rs: fleet_push a method, the raw point base + PUSH_FRAC (objective - base), moved front_w of the way to the front centre nearest it. meter.rs: lean(post, moved) once a second; leans by group in the census line.
Scenario form. "centre" alone as an anchor; [[tank]] toward; met_beyond = { by, on, at, more_than } over a new Outcome::hit_at.
Family. scenarios/families/a-raid-on-a-guarded-base.toml, 400 runs, floor 0.75. Sweeps: make scenario S=... RUNS=400 SET="w_front_push=0" and the others in section 61's table; the bar swept over 8 to 25 tiles at 200 seeds before 10 (BASE_R) was chosen.
Guard. build/lr61/run.sh: a frozen binary; the fair census on naval.map, 900 s, 3 respawns, seeds 0 to 47 in three jobs of 16; trial caps = baseline (the defaults), baseline the same with both weights
- The pooled difference is the mean of the 48 per-seed differences,
its error their standard deviation over the square root of 48.
IV. Results
| Measure | Before | After |
|---|---|---|
| The family, met beyond 10 tiles, 400 seeds | 50.5% +/- 2.5% | 85.5% +/- 1.8% |
| The first family (sight and drones), any bar from 12 to 25 | same with or without | same |
| Push point leaned, census seconds | 23,950 of 30,216 (79%) | |
| Guard post leaned, census seconds | 1,039 of 1,130 (92%) | |
| The guard, 48 seeds | -0.27 +/- 0.64 a seat | |
| Ashore dodges hit | 15.8% | 15.7% |
Tests: brain 82 passed and 11 ignored; client 32; net 14; proto 3; sim
- 100 weights, 17 bits; ten scenario files, six families with floors.
V. Discussion
- A raid seen is met far out already. The guard's shells and drones reach past where any post could stand; the post matters only when the guard is blind, and a scenario has to take the sight away to show it.
- The guard is seldom at its post. Since 17 October an idle guard rotates, so the push point is where its time goes, and the push point is the part that pays. The post leans for about 8 s a match.
- A part that acts and reads zero. Unlike LR 58's front read, the push point acts most of a rotating guard's time; the census still reads no gain or loss a point wide. Whether leaning helps a whole match is for the tuning phase.
For Paul: none blocking. Next: backlog 20 (island_site where a naval start chooses), then 48 (the starting numbers). Commits are local; nothing was pushed.