Development log · 2026-06-08

Soldier-Scale Realism — three felt wins, one honest revert

The player now watches this counterinsurgency sim at the soldier scale — and at that scale four things "didn't feel right." Each was turned into a number before it was touched. Two shipped as large measured wins, one shipped as a quieter one, and the fourth — the hardest — was built, measured against its own baseline, found to regress the most load-bearing system in the codebase, and reverted. The revert is in here on purpose: an honest logbook is part of the product.

Item 1 — Village dwell was ~70× too short, and the census teleported

shipped The complaint, as a number

A cordon-and-search is a half-day operation; a census enrolls a village person by person. In the sim, every on-station dwell was minutes — and the census flipped to "done" the very tick the squad arrived. A headless probe (scripts/dwell-probe.ts) read it off HEAD: census 300 s, with the "done" flag set at 0 s on arrival — a teleported state flip, not work.

mission	before	after	doctrine band
presence	180 s	3,600 s (1 h)	~1 h flag-showing
recon	150 s	4,500 s (1.25 h)	≥ 1 h
ambush	1,100 s	14,400 s (4 h)	hours of waiting
census	300 s, INSTANT	11,200 s, PROGRESSIVE	2–8 h, population-driven
cordon	300 s	7,700 s (2.1 h)	2–8 h half-day
overwatch	900 s	18,000 s (5 h)	3–8 h (24 h cap)
KLE / shura	360 s	5,856 s (1.6 h)	1–2 h per meeting

Why "×70 the numbers" would have been worse — and what was done instead

A static six-hour wait at 4× speed is more tedious than a five-minute one, not less. The fix is four coupled changes, not one knob:

• Population-driven duration. Census/cordon scale with village.population (biometric enrollment is up to 30 min per fighting-age person), clamped to a 2–8 h band. Every continuous on-station payoff (e.g. a shura's +8 attitude) was re-normalised by the dwell, so raising the duration doesn't secretly multiply the gains.
• Progressive census. A new censusProgress (0→1) climbs with time on station; "done" trips only at 1. Recall the squad early and the census is partial — and it persists, so a follow-up element pays only the remaining time.
• A dwell event-roll. The patient hours surface decisions: a weapons cache turns up under a false floor, an enrolled man pings the biometric watch-list, the Female-Engagement-Team gap blocks the women's quarters, a herder demands solatia for a shot goat (the Restrepo "Cow Incident"), a man bolts the cordon. Each is a modal the engine already knows how to pause time on.
• Warp is the intended way through. The skip-to-event warp loop already halts on a pending modal — so the player fast-forwards the patient hours and is yanked back the instant something matters. Long-but-real becomes tension, not a progress bar.

Item 2 — The squad sat on a base of fire while one man shuffled

shipped The complaint, as a number

The owner watched a firefight and saw the whole squad go to ground and stay there — at most a man or two repositioning, never a maneuver element. A probe (scripts/squad-maneuver-probe.ts) confirmed it on HEAD and found three damning facts: under the default Hold drill the squad never autonomously assaulted; the squad's standing order (SOP) was cosmetic — Hold, Suppress and Assault all spent ~45% of contact breaking contact and behaved almost identically; and even under an explicit Assault order, bounding discipline was 0% (the whole maneuver team rushed at once) along a route whose flank-cover ratio was 0.92 — i.e. it charged through less cover than a straight line. A frontal banzai rush into the open.

metric (heat 0.45)	Hold before→after	Suppress	Assault
enters an assault	0%	0%	46% (frontal)
enters an assault	→ 22%	→ 10%	→ 29% (bounding)
bounding discipline	—	—	0%
bounding discipline	→ 100%	→ 100%	→ 100%
flank-cover ratio (when flanking)	0.92 → 1.68–1.84  (route cover ÷ straight-line cover; >1 = a covered flank)

The fix gives the maneuver decision to the squad-leader AI — never to the player, who still only sets the SOP and approves fires (that is the game's soul: "combat is 100% AI; the hardest part of command is watching"). Four mechanisms:

• Develop-then-commit. React to contact = get down and return fire; a beat later, the moment the squad can gain fire superiority, the leader commits a maneuver fire team — unbidden.
• The SOP becomes a real lever via a develop-timer. An earlier attempt to key the lever on how-pinned-the-squad-is failed: that signal is bimodal (either winning or pinned), so it rarely lands between thresholds and Hold ≈ Assault. The robust fix is time: Assault commits at once, Hold develops ~7 s first, Suppress ~18 s and meanwhile calls fires — so many contacts resolve before the cautious orders ever assault. Now Assault > Hold > Suppress at every heat.
• A covered flank, not a beeline. The maneuver element routes to a flank objective off the enemy's axis via the cover-biased planner; frontal is the fallback only when no covered flank exists.
• Bounding overwatch. The element splits into buddy pairs and only one pair rushes at a time while the other overwatches — the single most recognisable infantry behavior, and it was entirely absent before.

Item 4 — The squad rings the spur instead of climbing it

tried · measured · reverted Why this section exists

This is the part most reports leave out. A squad ordered to a peak observation post planned a 3.2 km route around a 468 m climb — it ringed the spur instead of switchbacking up the face. The cause was clear and the fix was textbook (anisotropic walking cost so a traverse is cheaper than a straight climb, plus softening the hard cliff cutoff). It was built. And then the numbers said: don't ship it.

A fair probe (scripts/op-route-probe.ts) measured the detour against a fixed objective — the highest peak reachable under the strict cutoff, identical before and after — comparing the live tree to a clean checkout of HEAD in a git worktree:

attempt	mean detour	verdict
HEAD baseline	×3.78	the problem
anisotropic Tobler cost alone	×4.52	worse — over-penalises, pathological zig-zag
directional cost (right magnitude) + turn penalty	×3.13	better detour, +5pp reachability… but stalled a squad
REVERTED to baseline	×3.78	0 movement stalls — clean

The micro-terrain half of the same foundation — letting a soldier take cover behind that rock or low wall, not an averaged 5 m cover cell — is scoped alongside it as docs/issues/020; the flank work above already showed why it matters (a meaningfully-covered flank exists only 3–7% of the time on the coarse cover raster, even though the flank mechanism itself is proven).

How it was verified

• Every complaint became a number before any code changed — three purpose-built headless probes (dwell-probe, squad-maneuver-probe, op-route-probe), each baselined on HEAD first.
• Before/after on identical inputs. Where a change altered reachability (so the probe's own objective would move), the comparison was run against a clean checkout of HEAD in a git worktree, on a fixed objective selected by a mover-faithful flood that honours the planner's anti-corner-cut rule.
• Held-out seeds. The maneuver and OP wins were read on seeds that were never in the tuning loop, not just the ones tuned on.
• Live, in the deployed game. Both shipped items were driven in the real app via window.__ITM — the dwell modal rendered and resolved correctly; the squad was caught mid-flank and its element split dumped frame-by-frame.
• The standing gate held green: tsc, build, smoke (serialize round-trip), and balance (no stall) on every shipped change; new persisted state went into both serialize() and loadWorld() (a save-version bump to v7).

Engineering record: docs/progress/2026-06-07-soldier-scale-impl/ (baselines, after-numbers, live dumps). Deferred work: docs/issues/019 (any-angle OP planner), docs/issues/020 (micro-terrain cover objects). Probes: scripts/{dwell,squad-maneuver,op-route}-probe.ts. Built and verified by an autonomous agent; numbers are quoted as measured.