After-Action Report · Five-times-better campaign

The 5× Campaign Making a defense-grade war-sim READ, SOUND, and PLAY true

The simulation was already defense-grade under the hood; this campaign made it read, sound, and play true.

In the Mountains is a deterministic, single-clock real-time counterinsurgency sim of a remote US combat outpost in a Korengal-like valley. The engine — scored base-of-fire vs. maneuver, PID-aimed indirect fire, ambush/scoot insurgent AI, elevation LOS over a 512² terrain grid — was deep before this campaign began. What it lacked was legibility, sensory presence, a living strategic layer, and a true-to-scale map. Five file-disjoint waves closed those gaps, each gated by a headless harness and an independent adversarial verify. Below is every hypothesis, every number, every mechanism — and every residual we did not fully solve.

BRANCH main BASELINE @ c6bf324 WAVES A · B · C · D · E DONE-GATE tsc · build · smoke · balance

Headline scorecard — before → after

Combat exchange ratio

~5:1 against→US-favorable

balance.ts · KIA 0.75→0.33

Enemy accounted / contact

1.25→4.42

balance.ts · 3.5× · held-out 5.40

Suppression (reachable)

0.000→1.0

realism-probe · 27–36% pinned

Enemy elevation adv.

+2.41 m→+38 m

plunging fire (ambush)

Stance: standing under fire

73%→~10%

prone 0%→1.7–4% · crouch ~87%

Audio cue fidelity

silent→1:1

audio-probe · 1252==1252, 0 dropped

COIN tour-score spread

0.5→30.3

campaign-loop · careful vs body-count

CERP projects completed

0 (sabotaged)→1.3–2.0

4/4 on survey-40

Soldier figure vs. true size

36×→17× → 3×

scale-probe · default→tactical zoom

Squad reads as

one blob→9 spaced men

overlap true→false at every zoom

Every figure above is reproduced verbatim from the dated source docs in this folder. Numbers-first is a hard rule on this repo — nothing here is rounded up.

01The Method — orchestration as the differentiator

The work was not "do a big todo list." It was an orchestration: a recon fan-out to map the opportunity space, then file-disjoint waves run as pipelines, each ending in an independent adversarial verify — all gated by no fix without a number, the determinism contract, and a DO-NOT-RETRY list so no agent re-walks a dead end.

1 · Recon fan-out

12 domain-expert agents (infantry doctrine, ballistics, COIN, audio, render, game-feel…) read the engine in parallel and proposed improvements along realism and joy/awe. Their output was judged and deduped into 63 ranked hypotheses — each a falsifiable claim with a measurable target, not a vibe.

2 · DO-NOT-RETRY ledger

Previously-reverted approaches travel with their measured failure so they are never re-proposed: screen-shake-as-shockwave / recoil-kick (reverted, fought the camera pose), a WeakMap for civilian home (rejected — breaks serialize-replay determinism), and coarse-dt combat scoring (caught lying: 23 KIA at DT=30 vs 11–12 at fine-dt on the same seed).

The recon → wave → verify pipeline

Read left-to-right. Each wave is a self-contained pipeline; the lanes run in parallel because they own disjoint files. Nothing is "done" until a separate, skeptical agent re-derives the number. (The diagram shows the original four lanes; a fifth lane — Wave E, world-map scale — landed after the same recon, on the render layer, and followed the identical metricize → baseline → re-measure → adversarial-verify shape.)

Pipeline. One recon fan-out feeds four file-disjoint wave lanes. Each lane metricizes on a headless harness, baselines on HEAD, implements, then hands to a fresh verifier who re-derives every number before merge. The verify column is where the work is proven — not where the author says "done."

File-partition map — why the waves could not collide

Parallel agents are safe only with one writer per file. The partition below is the contract: each wave's owned files are disjoint, so the four ran concurrently and merged cleanly. The architecture's strict layering (pure engine → React-free renderers → the single Zustand bridge) made the cut-lines natural.

WAVE Aengine, deterministic — lib/sim/combat.ts · weapons.ts · ballistics.ts · ai/friendly.ts · ai/insurgent.ts · world/director.ts

WAVE Bnew package + render bridge — new lib/audio/{cue,mapper,synth,player,index}.ts · state/store.ts⁽⁺⁾ · components/world/WorldView.tsx⁽⁺⁾ · scripts/audio-probe.ts

WAVE Crenderers, React-free — lib/render/{topo,combat-fx,draw}.ts · lib/sim/ai/civilian.ts · components/world/WorldView.tsx⁽⁺⁾ · scripts/atmospherics-probe.ts

WAVE Dstrategic engine + HUD — lib/sim/world/{world,types,create,projects,tasks,events}.ts · new directives.ts · campaign.ts · scripts/campaign-loop.ts · state/store.ts⁽⁺⁾ · DeployScreen.tsx

WAVE Erender layer, React-free — lib/render/draw.ts · lib/render/combat-fx.ts (import-sync only) · components/world/WorldView.tsx⁽⁺⁾ · scripts/scale-probe.mjs

⁽⁺⁾ Shared touch-points (store.ts, WorldView.tsx) were reconciled at merge — the one place a cross-wave regression appeared (the Space-key collision, §04) and was caught by the verifier.

The five laws every wave obeyed 1. No fix without a number — baseline on HEAD first, re-measure each change in isolation. 2. Don't trust yourself — a separate skeptical agent re-derives the result. 3. Prove on a held-out tail — tune on survey-0..39, prove on survey-40..59. 4. The oracle obeys the mover's real rules — a loose oracle lies as badly as the metric. 5. Numbers-first, residual named, restraint logged — lead with the unflattering figure.

02Wave A — Combat Soul · the firefight now reads true

Hypothesis — the exchange ratio was inverted

On HEAD, a 12-deployment balance run (balance.ts 12 50) produced 0.75 US KIA and 5.58 US WIA against only 1.25 enemy accounted per contact — roughly 6.3 friendly casualties for every 1 enemy, an inverted ~5:1 ratio. Real US infantry won firefights tactically; the sim had it backwards. Two candidate causes: under-effective US fire, over-lethal enemy fire, or an undercount of fighters who exfil/evac. The deeper realism reads were equally damning: 0.000 suppression (no one was ever pinned), 73% of soldiers stood upright under fire (0% prone), and the enemy held a meaningless +2.41 m elevation edge in a valley defined by plunging fire.

Method — the headless harness

scripts/balance.ts <deployments> <minutes> runs continuous deployments and reports KIA / WIA / enemy-accounted / civ-cas plus a stall check. Realism reads came from realism-probe.ts (suppression reachable/pinned fraction, stance distribution, terminal ballistics, elevation advantage, shoot-and-scoot displacement, enemy massing). HEAD numbers were captured verbatim before any edit, then every change re-measured in isolation, then re-proven on a fresh 20-seed held-out tail.

Change — six engine files, the right way

Touched combat.ts, weapons.ts, ballistics.ts, ai/friendly.ts, ai/insurgent.ts, world/director.ts. The fix was not a damage buff that turns the game into a turkey-shoot. Instead: suppression now credits the whole flight path of a round (everyone near the line gets pinned, not just the target), soldiers take cover and go prone/crouch under fire, ambushers occupy real high ground, and they shoot-and-scoot after engaging. US fire became effective because the tactics became real — KIA fell while enemy accounted rose.

Verified by the adversary The Wave-A verifier reran balance.ts twice → byte-identical (determinism intact), confirmed the civClear civilian-fire gate untouched, and found no DO-NOT-RETRY violation. Civ casualties stayed 0 throughout.

Result — before → after (with held-out)

Source: results-waves-ab.md · `balance.ts 12 50`; held-out 20×50 fresh seeds.
balance.ts metric	HEAD	After (12×50)	Held-out (20×50)
US KIA / contact	0.75	0.33	0.30
US WIA / contact	5.58	3.92	2.70
Enemy accounted / contact	1.25	4.42 (3.5×)	5.40
Exchange ratio	~5:1 against	~even	US-favorable
Civ casualties	0	0	0
Stalls	none	none	none

The settled mid-campaign measurement (a separate balance pass cited in the baseline lineage) landed at 0.58 KIA / 4.33 WIA / 5.67 enemy — the chart in §07 plots all three after-states against HEAD so you can see the convergence, not just one cherry-picked column.

Realism wins (realism-probe)

Source: results-waves-ab.md · realism-probe.ts.
#	Metric	HEAD	After
1	Suppression reachable / pinned	0.000 / ~0	1.0 / 27–36%
2	Stance standing / prone	73% / 0%	~10% / 1.7–4% (crouch ~87%)
5/18	5.56 kill% @25 m / terminal dmg	0.7% / 30.7	2.1% / 42.2
9	Enemy mean elevation adv.	+2.41 m	+38 m (ambush)
4	Shoot-and-scoot displacement	0.02%	~81% of engage entries
11	Frags (grenade & clear)	dead code	wired live (civ-gated ≤35 m)
12	Enemy within-40 m massing	0.00	1.4–2.1 (L firing line)
30	Night tempo	inverted (1.67× at night)	flipped (deep-night ×1.8 delay)
—	Wind drift (bonus)	flat ~0.23 m	0.31 → 2.45 m w/ range

How it works — suppression credit along the whole flight path

A fired round is not a point event at the target; it traces a line from muzzle to impact.
Every soldier within a band of that line accrues suppression — the rounds "snapping past" are what pins real infantry, not just the rounds that hit. This is why HEAD's 0.000 (suppression only credited on a hit) jumped to a 1.0 reachable ceiling with 27–36% pinned at any instant.
A suppressed soldier's behaviour changes: he crouches or goes prone and his fire degrades — which is exactly why standing-under-fire collapsed 73%→~10%. Suppression is the mechanism that makes the firefight feel like a firefight, and the reason US KIA fell: pinned enemies shoot worse.

DAY — combat-day.pngNIGHT — night-firefight.png

Daytime firefight on the green valley floor

Day. Full ambient light: the green valley floor, vegetated river, and HESCO cross-hatch are all legible. 1st squad (Kowalski · Beckett · Carrillo · Salazar · Whitaker) now reads as distinct, evenly-spaced individuals — the Wave-E scale fix (§06), not a single 21 m giant. The same scene at night is the right-hand frame — the day/night pair below.

Night firefight with additive muzzle and blast light

Night. The night wash darkens the valley; additive warm glow from muzzle, blast, and tracer (Wave C, gated to night>0.25, never cyan) paints the firefight on top — the single most cinematic change in the campaign.

Honest residual (Wave A) Suppression element-average peak (0.43–0.48) sits below the guessed 0.6–0.85 band: the per-contact mechanism is decisive, but fights are short/sharp and pushing harder risks the KIA guardrail (~35% confidence it converges without breaking balance — logged, not forced). Mean enemy pairwise spacing is not reliably down; elevation (+38 m) sits above the guessed band but is deemed more Korengal-faithful, not less.

03Wave B — Procedural Audio · the absent transformative dimension

Hypothesis — the game was silent

The repo had zero audio dependencies and the game made no sound. A firefight you can see but not hear loses most of its weight. The constraint was hard: audio must not touch lib/sim (determinism contract) and must add no binary assets.

Method — a 1:1 cue oracle

scripts/audio-probe.ts asserts a strict invariant: exactly one audio cue per sim effect-id. It replays the deterministic sim, counts non-blood effect ids, and counts cues — they must match with zero dropped and zero double-fired. It also replays twice and demands the cue stream be byte-identical (the mapper must be pure).

Change — a new `lib/audio/` package on the render seam

Five new files (cue / mapper / synth / player / index) plus wiring in state/store.ts, WorldView.tsx, and a HUD volume/mute control. The render-side event seam is the key: the audio system subscribes from the RAF frame() loop the same way the visual effects do — it reads sim.effects (muzzle / impact / ricochet / blast / blood / frag-air / smoke-pop / flare), the LogEntry bus, and FireMission status — and never imports into the engine.

How it works — the render-side audio event-seam

Pure mapper (mapper.ts): consumes the sim's Effect / LogEntry / FireMission stream and emits AudioCue[]. It is headless-testable and deterministic — two replays produce byte-identical cues (7394 cues verified). It contains no Web-Audio calls, so it can run under tsx.
Browser synth (synth.ts / player.ts): Web-Audio oscillators + filtered noise + envelopes. Distance attenuation, a low-pass on far sounds, a speed-of-sound delay, stereo pan and zoom scaling — all procedural, no samples.
Determinism firewall: lib/sim has zero audio imports. The AudioContext unlocks only on a user gesture (Next 16 App Router + browser autoplay policy), and audio is silent during pause/warp. The engine never knows audio exists.

Result — verified invariants

Source: results-waves-ab.md · `AUDIO OK`.
Invariant	Result	Verdict
Cue-per-effect fidelity	1252 cues == 1252 effect ids · 0 dropped · 0 double-fired (of 1480)	1:1
Mapper purity / determinism	two replays byte-identical · 7394 cues	pure
Determinism contract	`lib/sim` has ZERO audio imports	intact
Voice bounding	24-voice pool cap	bounded
Autoplay policy	unlocks only on user gesture	compliant
Pause / warp	silent during pause and time-warp	gated

Final sound-quality judgement requires a human at a browser — that capture is a real-app pass, not a headless one. The invariants above prove the plumbing is correct and determinism-safe; they cannot prove a 7.62 sounds menacing. That is the one honest caveat for Wave B.

04Wave C — Sensory & Game-Feel · the valley sleeps, wakes, and weathers

Hypothesis — the valley was diurnally flat and weatherless on screen

Two sensory gaps. First, civilians wandered 24/7 at a flat 100% outdoor occupancy with no dawn/dusk rhythm, and the flagship COIN tell — civilians melting away before an ambush, which the tutorial explicitly promises — did not exist (0% drop). Second, weather existed in world.state but was never rendered; night was a flat wash with no light from the fight.

Method — the atmospherics probe with a mover-faithful oracle

scripts/atmospherics-probe.ts measures outdoor occupancy by hour, and a staged pre-contact melt-away, using ground-truth time from world.secondsOfDay (independent of the sim.light the brain reads) and "home" = the same snapped point the mover actually walks to. Law 4 in action: the oracle obeys the mover's real rules, so it cannot be fooled by the system it checks.

Change — render-side atmospherics + civilian brain

Owned lib/render/{topo,combat-fx,draw}.ts, lib/sim/ai/civilian.ts, and WorldView.tsx. The diurnal pattern and melt-away are deterministic engine changes in the civilian brain (no new persisted field — serialize stays byte-identical). Weather rendering, night-light, wind-leaned smoke, camera-punch, and the off-screen contact pointer are render-only.

Result — headless-verified (deterministic)

Source: results-wave-c.md + issue 016 · atmospherics-probe.ts.
Target	Result	Verdict
A · Diurnal pattern-of-life	day 80–98% out → 19:00 10% → night 0% → 06:00 61%; survey-43 held-out night 14.8%/23.9%, midday 62%	PASS
B · Pre-contact "calm before" melt	survey-7: 56% of outdoor civs clear before a staged ambush, 0 shots, children first; control 0%. Held-out 48%, kids-first t=42 s	PASS
Determinism (A+B)	same-seed civ-position hashes byte-identical in- & cross-process; smoke round-trip ×3	PASS

Code-verified, capture in the real app

Source: results-wave-c.md — code-verified; live screenshots below.
Target	What changed	Verdict
C · Weather rendered	`topo.ts drawWeather()`: rain streaks / drifting fog / elev-gated snow-line / blowing dust from `world.state.weather` + live wind vector	CODE
D · Night light	`combat-fx.ts drawNightLights()`: additive warm glow from muzzle/blast/tracer, gated night>0.25, no cyan, age-faded	CODE
E · Wind-leaned smoke/haze	`draw.ts drawSmoke()` + `combat-fx.ts drawCombatHaze()` drift with the wind vector	CODE
F · Camera-punch + edge-flash	`noteShakeEvents()/drawEdgeFlash()` — capped 7 px, prox²·size, monotonic-id dedup; not the reverted shockwave/recoil	CODE
G · Off-screen contact pointer	`drawOffscreenContactPointer()` + jump-to-contact key snap (ppm ≥ 1.2)	CODE

How it works — night-light additive over the night wash

At night the renderer lays a darkening night wash over the valley.
drawNightLights() then paints additive warm glow from each live muzzle, blast, and tracer effect on top of that wash — additive blending means light adds brightness rather than replacing colour, so a firefight genuinely illuminates its surroundings.
It is gated to night > 0.25 (no glow in daylight), the glow is age-faded with the effect's life, and it is deliberately warm — never cyan — so it reads as fire, not UI.

Cross-wave regression — caught by the verifier (the value of the skeptical pass) WorldView.tsx bound jump-to-contact to Space, but DeployScreen.tsx already bound Space → pause; both window listeners fired on one press (Space paused and jumped the camera). Fix: rebind jump-to-contact to 'c' only. A green author-run would have shipped this — the independent adversarial pass found it. This is precisely why "don't trust yourself" is Law 2.

WEATHER — rainWEATHER — fog

Rain. drawWeather() draws streaks leaned along the live wind vector; the floor desaturates under overcast. Combat FX remain readable through the weather layer.

Fog. Drifting fog reduces apparent visibility and mood without hiding the navigational anchors. Snow-line is elevation-gated (not shown at this floor altitude).

05Wave D — COIN Soul · "counterinsurgency is the real game" made true

Hypothesis — the strategic pillar was mechanically dead

DESIGN pillar 5 is "you can win every firefight and still lose the valley" — yet on HEAD, playing COIN well vs. badly produced near-identical outcomes. campaign-loop.ts showed careful-COIN tour score 38.0 vs. body-count 37.5 — a spread of 0.5 (it needed >25 to discriminate). Worse: 0/6 projects ever completed (all sabotaged because no element ever stayed to hold the multi-day build), CERP was a one-way drain to 0, directive deadlines/penalties were written but read nowhere, and 5 of 7 directive kinds were never even created. The soul of the game was inert.

Method — the campaign-loop harness, two scripted policies

scripts/campaign-loop.ts drives full multi-day tours under two scripted commander policies — careful COIN vs. body-count — and measures whether the strategy layer discriminates: tour-score spread, mean-attitude swing, projects completed, CERP two-way, directive cadence + deadline-fire. A wide spread means the game rewards patience; a flat one means the pillar is dead.

Change — secure-build, two-way CERP, live directives, asymmetric promises

Engine files world / types / create / projects / tasks / events + new directives.ts, plus campaign.ts, the harness, and HUD wiring in store.ts / DeployScreen.tsx. The keystone is the secure-build order: a patrol-level garrison order (World.secureBuild(memberIds, villageId, …)) routes an element to a project site (reachability-aware, no beeline) and holds an open-ended overwatch — so securityAt finally counts a held element through the whole build, and projects complete instead of sabotage.

How it works — the secure-build order that makes projects complete

A multi-day CERP project (a micro-hydro, a school) has a building stage that needs sustained security. On HEAD, securityAt counted any 2 transient friendlies — so a patrol passing through "secured" it for a tick, then left, and the build sabotaged.
secureBuild issues a patrol a secureHold task (tasks.ts): the element routes to the site reachability-aware and holds an open-ended overwatch until the build finishes, sabotages, or it's recalled. securityAt now counts the held element first.
A completed project pays off: per-type PROJECT_PAYOFF (×1.6 if the village wanted it, ×0.6 if not) + a CERP refund + a fulfilled-ask hook. Winning the valley now out-weighs the body count in computeTourScore (coin = projComplete×11 + dirComplete×4 + kept×3 + wonOver×3; attrition only ×0.06).

Result — the flip (default 3×8 tour)

Source: results-wave-d.md + issue 015 · `campaign-loop.ts` 3×8.
Metric	HEAD	After	Verdict
Tour score: careful vs body-count	38.0 vs 37.5 (0.5)	60.0 vs 29.7 (30.3)	discriminates
Careful COIN moves mean attitude	Δ +0.3 (flat)	Δ +7.7 (body-count +0.1)	PASS
CERP projects completed / tour	0 (all sabotaged)	1.3–2.0 (4/4 survey-40)	PASS
CERP economy	one-way 30000→0	two-way (stipend + refund)	PASS
Directives	dead; never fail	cadence 2.3–3.3/tour; fail FIRES (D14 presence −10)	PASS
KLE / shura	360 s attitude-drip	kept +10 / broken −12 (asymmetric)	PASS
Secure-build order	none (#1 API gap)	World.secureBuild()	PASS

Determinism — the #1 risk of this wave, independently confirmed Every new persisted field is in both serialize() and loadWorld() (nextCerpStipendAt, nextDirectiveAt, civCasualties, reliefWatchClock; village ask/keptPromises/brokenPromises; Directive.startMetric; Task.secureVillageId). The proof: a serialize → load → serialize round-trip is BYTE-IDENTICAL (240,893 == 240,893) — the strongest possible evidence no field was dropped from load.

The command layer — coin-hud.png

The HUD is where the strategic layer becomes playable: the directive feed, SOP settings, fire-support, and logistics are all surfaced so a human can drive the levers the harness drives. Callouts below point at the real panels.

The COIN command HUD with directives, SOP, fire support and logistics

COIN command HUD (COP Vimoto, fivex-hud). 1 the live battalion directives feed with status & deadline (here "Establish Presence", 12 days); 2 the intel / combat / command logs (CERP micro-hydro funded at Chichal); 3 mission orders (PATROL · RECON · INTERDICTION · CENSUS · SEARCH · ESTABLISH OP) and the SOP (pace + ROE that locks in contact); 4 the task organization roster (9/9 ready, SL/TL/SAW/GRN/RFL); 5 the logistics bars (5.56 / 7.62 / 60 / 81 mm, build, med kits, water); 6 the fire-support rail with the approve/deny levers that are the only in-combat command.

Honest residual (Wave D) — characterized, not closed → issue 015 The held-out 4×14 tail does NOT discriminate on average (careful 27.5 vs body-count 30.5, spread −3.0) — but this went from inverted (−26.8) to bimodal: where a careful tour survives it CRUSHES body-count (survey-40 careful 96, 4/4 projects, vs 29 — the seed that scored 2 before the fix), while on 2/4 combat-heavy seeds a scripted presence patrol routed at the most hostile village (adversarial) takes a catastrophic firefight → early relief → 0. Root cause: coarse-dt harness + tactically-naive scripted routing, NOT the COIN engine (which completes projects and moves attitude Δ+7.6 on every held-out seed). The validate stage fixed the instrument (adaptive fine-dt sub-step while a mobile patrol is in contact) rather than curve-fitting the score, and a process win lived: it caught its predecessor's mislabeled "all-PASS" and traced it to the harness over-killing (23 KIA at DT=30 vs 11–12 at fine-dt). Law 4 again.

06Wave E — World-Map Scale · the squad stops being a blob

Hypothesis — the one clearly-visible realism gap left on the map

Every other wave fixed something you had to measure to see. This one you could see at a glance and never un-see: soldiers were drawn ~36× too large. A fixed 15 px figure floor rendered a man at roughly 21 m of ground-width at the default zoom — a single soldier as wide as a nine-man squad's entire correct 5.5 m dispersion. So a squad collapsed into one overlapping blob at every zoom level, the opposite of what a defense-grade map should show. The fix is render-only (issue 014 "Bucket 1" = R1 + R2); the terrain/sim scale (Bucket 2) was deliberately left untouched.

Method — a zero-dep scale probe with a true-footprint oracle

scripts/scale-probe.mjs (seed korengal) sweeps the camera across zoom bands (pixels-per-metre 0.3 → 8.0) and at each band computes three independent ground-truths: figVsTrue (drawn figure width ÷ a real 0.6 m man), soldier/COP (the figure's footprint as a fraction of the 170 m COP — a realistic soldier is ≈1%), and squadFiguresOverlap (do two adjacent men in a 5.5 m-spaced squad draw on top of each other?). HEAD was captured verbatim before any edit; every band re-measured after; smoke round-trip re-asserted.

Change — figures track their footprint, squads aggregate into NATO icons

Two owned render files (lib/render/draw.ts, plus an import-sync of combat-fx.ts) and the WorldView.tsx bridge. The figure-size formula was rebuilt to track the real kit footprint instead of a giant legibility floor, a per-squad NATO icon was added below tactical zoom, and — the highest-risk detail — combat-fx was made to import the same figurePx so every cue stays glued to the now-correctly-sized men.

How it works — footprint-tracking size · legibility floor · per-squad LOD · one source of truth

R1 — the figure tracks its footprint. figurePx went from max(15, min(40, ppm·7)) to max(7, min(26, ppm·1.6)) — a 1.6 m kit footprint with a small 7 px legibility floor instead of a 15 px giant floor (dotR likewise max(4.5,min(13,0.95·ppm)) → max(3,min(9,0.7·ppm))). The LOD fade bands were pushed up to tactical zoom (FIG_FADE0 0.5→2.5, FIG_FADE1 0.9→3.5): individuals only resolve at ppm ≥ 2.5, exactly where they are near-true size and fit the 5.5 m gap.
R2 — per-squad aggregation below tactical zoom. A new drawSquadIcon() paints one NATO friendly rectangle (echelon dot + a "1st ×9" badge, contact-tint) per squadId at the element centroid; WorldView groups live non-garrison friendlies by squad below FIG_FADE0 with a 0.4-ppm crossfade so nothing pops. A squad reads as one icon zoomed out, as spaced individuals zoomed in — the right object at the right scale.
combat-fx SYNC — the #1 regression risk. combat-fx had its own private figurePx and inline dot formulas; those were deleted and it now imports figurePx / dotR from draw.ts (no import cycle). One source of truth means the suppression crescent, the bleed pool, and the danger-close halo can never detach from a resized figure ring — exactly the kind of split-brain that a naive size change would have introduced.

Result — before → after (scale-probe, seed korengal)

Source: results-wave-e.md · `scale-probe.mjs` (seed korengal). figVsTrue = drawn figure ÷ a 0.6 m man (lower is truer); soldier/COP = footprint ÷ 170 m COP (≈1% is realistic); overlap = two 5.5 m-spaced men draw on top of each other.
ppm (zoom)	LOD	figVsTrue	soldier/COP	squad overlap
0.3 (strategic)	squad-icon	83× → 39×	0.294 → 0.137	true → false
0.7 (default)	squad-icon	36× → 17×	0.126 → 0.059	true → false
1.0	squad-icon	25× → 12×	0.088 → 0.041	true → false
2.0	squad-icon	13× → 6×	0.044 → 0.021	true → false
4.0 (tactical)	figures	12× → 3×	0.041 → 0.010 (~1%)	true → false
8.0	figures	8× → 3×	0.029 → 0.009	true → false

squadFiguresOverlap is now false at every zoom: below 2.5 ppm a squad is drawn as one NATO icon (so there is nothing to overlap), and at 2.5+ ppm each figure (≤1.8 m) finally fits inside the real 5.5 m gap. At tactical zoom soldier/COP lands at ~0.010 (~1%) — a realistic footprint — down from a 36× giant at default.

scale-probe (seed korengal) · figVsTrue collapses across every zoom band — 36× → 17× at the default zoom, 12× → 3× at tactical, where a soldier finally reads at ≈1% of the COP. The teal dashed line is true scale (1×); below 2.5 ppm the squad is a single NATO icon (so figVsTrue describes the icon's stand-in), at 2.5+ ppm it is individual men. squadFiguresOverlap is false in every band.

Verified by the adversary + the done-gate The fix is render-only — lib/sim is untouched, so the determinism contract is intact by construction. The standing done-gate is green: tsc clean · build 0 errors · smoke OK, and the re-captured combat shots (this report's §02 day/night and §08 tableau, ppm 4.6) were inspected to confirm the combat-FX cues stay hugged to the resized figures — the exact regression the import-sync was designed to prevent.

Honest tradeoff / scope (Wave E) The legibility-vs-realism tension lives only in the narrow 2.1–2.5 ppm crossfade band (the icon fading out while individuals fade in — both briefly semi-transparent); it is unambiguous above 2.5 and below 2.1. R3 (render a village as a sub-compound cluster — a hamlet, not one qalat) was deliberately deferred as the smallest slice. Bucket 2 (valley shape, weapons-squad count, COP geometry) was left untouched — the sim/terrain is sensitive and issue 014 itself rates the relief and peaks as already authentic. Net: issue 014 → Bucket 1 RESOLVED; Bucket 2 + R3 deferred.

07The Charts — every headline number, drawn

Inline SVG, zero dependencies (matching the repo's ethos). Every axis is labeled and every chart cites its harness. Bars are colour-coded: ■ HEAD (before) · ■ after (tuned) · ■ after (held-out).

balance.ts 12×50 · HEAD: 0.75 KIA / 5.58 WIA / 1.25 enemy → settled 0.58 / 4.33 / 5.67 · wave-A 0.33 / 3.92 / 4.42 · held-out (20×50) 0.30 / 2.70 / 5.40. Three after-states are shown so the convergence is visible, not a single cherry-picked column. The inversion is gone.

realism-probe · suppression went from 0.000 reachable (never credited) to a 1.0 ceiling, with 27–36% of an element pinned at any instant. The element-average peak (0.43–0.48) is the logged residual — see §09.

realism-probe · the enemy's elevation edge in ambush jumped from a meaningless +2.41 m to +38 m — real plunging fire from the draws, faithful to the Korengal terrain.

realism-probe · standing-under-fire collapsed from 73% → ~10%; prone rose from 0% → 1.7–4%; the rest crouch (~87%). Soldiers now take cover like soldiers.

campaign-loop 3×8 · careful-COIN vs body-count went from 38.0 / 37.5 (spread 0.5) to 60.0 / 29.7 (spread 30.3). Patience now wins the valley — the design pillar is mechanically alive. The held-out 4×14 residual (spread −3.0, bimodal) is in §09.

audio-probe · 1252 cues == 1252 non-blood effect ids, 0 dropped, 0 double-fired (of 1480 total cues). The mapper is pure (two replays byte-identical, 7394 cues) and lib/sim has zero audio imports — determinism intact.

atmospherics-probe (survey-7) · HEAD outdoor occupancy was a flat 100% all 24 hours (day == night). After, the valley wakes at 06:00 (61%), peaks at midday (83%), empties by 19:00 (10%), and sleeps at night (0%). Held-out survey-43 night floor (≈13%) is real civilians fleeing infiltrators after dark, not a bug.

08Annotated Screenshots — what you are looking at

The featured screenshots are annotated inline at their point of use (combat day/night in §02, weather in §04, the COIN HUD in §05). All combat shots were re-captured after the Wave-E scale fix (§06) — the squad now reads as small, evenly-spaced individuals, not a giant. The remaining tableau below shows combat legibility — the Combat Visual Language reading clearly in daylight, the proof that the firefight is not just simulated but readable.

Daytime combat tableau showing the combat visual language

Combat tableau (day). The Combat Visual Language reads at a glance: 1 indirect-fire beaten zones bloom where mortars land; 2 the COP and helipad anchor the center; 3 the squad is a row of small, spaced named men (Pratt · Patterson · Marsh · Beckett) — the Wave-E scale fix (§06), each figure finally fitting inside the real 5.5 m dispersion; 4 the dark scree band is impassable terrain that dictates lines of sight; 5 the COP interior is modelled and labelled (TOC, motor pool, DFAC). Legibility is what lets a player read the fight they cannot micro-command.

09Honest Residuals — measured, characterized, logged

Honesty is a feature on this repo, not a weakness. Every item below is a known gap that was measured, root-caused, and logged for a future session — none is hand-waved, and none is rounded up to "fixed."

Wave A — suppression element-average peak below the guessed band

Peak 0.43–0.48 vs the guessed 0.6–0.85. The per-contact mechanism is decisive and the reachable ceiling is 1.0; fights are short/sharp, and pushing suppression harder risks the KIA guardrail. Calibrated ~35% confidence it converges without breaking balance — logged, not forced. Related: mean enemy pairwise spacing is not reliably down; elevation (+38 m) is above the guessed band but judged more Korengal-faithful, not a defect.

Wave B — sound quality needs a human in a browser

The headless harness proves the cue plumbing is 1:1 and deterministic, the mapper is pure, and lib/sim has zero audio imports — but it cannot prove a 7.62 sounds menacing or that the mix is balanced. Final sound-quality judgement is an explicit real-app capture, not a headless claim.

Wave C — weather is render-only; gameplay consequence deferred

Rain, fog, snow-line, and dust are drawn from world.state.weather, and they drift with the live wind vector — but they do not yet change LOS, accuracy, or movement. That is a deliberate scoping call: the visible layer landed first; weather-as-a-gameplay-modifier is a future wave.

Wave D — held-out 4×14 bimodal spread −3.0 (a harness artifact, not the engine)

The held-out tail does not discriminate on average (careful 27.5 vs body-count 30.5, spread −3.0) — but it went from inverted (−26.8) to bimodal: where a careful tour survives it crushes body-count (survey-40 careful 96, 4/4 projects). The 2/4 failing seeds are a coarse-dt harness + tactically-naive scripted patrol routing (presence patrols sent at the most hostile village → catastrophic firefight → early relief → 0), not the COIN engine — which completes projects and moves attitude Δ+7.6 on every seed. A clean held-out proof needs fine-dt patrol windows or smarter scripted routing. Logged in issue 015 for a future wave rather than chased with score curve-fitting.

Wave E — world-map scale: Bucket 1 RESOLVED; Bucket 2 + R3 deferred (issue 014)

The clearly-visible half of issue 014 is fixed: soldiers no longer paint as giants (figVsTrue 36× → 17× at default, 12× → 3× at tactical), a squad no longer blobs (squadFiguresOverlap true → false at every zoom), and a soldier's footprint lands at a realistic ~1% of the COP at tactical zoom — see §06. What remains deliberately deferred: R3 (render a village as a sub-compound hamlet rather than one qalat — the smallest slice) and Bucket 2 (valley shape, weapons-squad count, COP geometry — sim/terrain is sensitive and 014 already rates the relief/peaks as authentic). The narrow 2.1–2.5 ppm crossfade band is the one residual seam, unambiguous on either side of it. Honestly partial — but the part you could see with your eyes is closed.

The DO-NOT-RETRY ledger (negatives honoured) Three approaches were refuted earlier and are not re-attempted: screen-shake-as-shockwave / recoil-kick (fought the camera pose — only the capped 7 px camera-punch + rust edge-flash survived), a WeakMap for civilian home state (breaks serialize-replay determinism), and coarse-dt combat scoring (caught lying: 2× the KIA). Honour recorded negatives — never re-walk a measured dead end.

10The 5× Assessment — how much better, axis by axis

Is the game five times better? "5×" is a direction, not a literal multiplier — so the honest answer is per-axis and tied to a number or a screenshot. The verdict: the engine was always deep; this campaign converted that depth into felt experience — and on the two axes that were mechanically dead (the exchange ratio, the COIN pillar) plus the one that was visibly broken (the 36× soldier giant), the improvement is not incremental, it is qualitative. Below, each axis carries its evidence.

Realism

Combat lethality

The exchange ratio was inverted (~5:1 against, 0.75 KIA / 1.25 enemy). It is now US-favorable (0.33 KIA / 4.42 enemy, held-out 0.30 / 5.40) — and it got there the right way (suppression + cover + tactics), with KIA falling, not via a damage buff. A qualitative flip, not a tweak.

Enemy & terrain

Suppression 0.000 → 1.0 reachable; stance 73% → ~10% standing; elevation +2.41 → +38 m plunging fire; shoot-and-scoot 0.02% → ~81%; enemy massing 0.00 → 1.4–2.1 (an L-shaped firing line). A soldier reading FM 3-21.8 would now recognize the behaviour. (Residual: suppression element-avg peak, §09.)

Map scale & legibility

A squad used to draw as a single 21 m giant (a soldier 36× true size at the default zoom), the most visible un-realism on the screen. It now reads as nine spaced men at a realistic ~1% of the COP footprint (figVsTrue 36× → 17× → 3×; squadFiguresOverlap false at every zoom), and aggregates into a proper per-squad NATO icon when zoomed out — the right object at the right scale. (Residual: Bucket 2 sim/terrain scale + R3 village cluster deferred, §09.)

COIN strategy

The design pillar was mechanically dead (spread 0.5, 0 projects complete, dead directives, one-way CERP). It now discriminates (spread 30.3, attitude Δ+7.7, projects complete, live directives with firing deadlines, asymmetric promises). "Win every firefight and still lose the valley" is now playable. (Residual: the held-out harness artifact, §09 — the engine is sound; the instrument needs fine-dt patrol windows.)

Joy & awe

Audio

From silent to a deterministic, procedural Web-Audio layer with 1:1 cue fidelity (1252==1252), distance attenuation, speed-of-sound delay, and stereo pan — zero binary assets, zero engine coupling. The single most transformative missing dimension, added cleanly. (Caveat: final mix quality is a human-in-browser judgement.)

Night-light

The most cinematic change in the campaign: additive warm muzzle/blast/tracer glow over the night wash (drawNightLights(), gated night>0.25, never cyan). See the night-firefight screenshot in §02 — the fight genuinely lights the valley.

Weather & atmospherics

Weather went from data-only to rendered — rain streaks, drifting fog, elev-gated snow-line, blowing dust, all leaned along the live wind vector (§04 screenshots). The valley now reads as a place with weather. (Honest: it's render-only; no gameplay consequence yet, §09.)

Living world (pattern-of-life)

The valley was diurnally flat (100% outdoors, 24/7) and the flagship COIN tell — civilians melting away before an ambush — did not exist. Now the valley sleeps and wakes (0% night → 83% midday) and civilians melt away before a shot is fired (56% clear, children first, 0 shots). A player can finally read the absence the tutorial promised.

Game-feel

Capped camera-punch (7 px, prox²·size), a rust danger-close edge-flash, wind-leaned smoke/haze, and an off-screen contact pointer with jump-to-contact — the allowed felt-weight mechanics, not the reverted shockwave/recoil. The fight you cannot micro-command now tugs at the camera when it matters.

Strategic stakes

The command HUD (§05) surfaces directives, SOP, fire-support, task-org, and logistics — the levers that were previously inert are now visible and drivable. Combined with the live COIN layer, the campaign now has real stakes: a tour can be won by patience and lost by impatience, independent of the body count.

The bottom line Before this campaign a skeptic could play and not realize how deep the engine ran — it looked flat, sounded silent, and rewarded the wrong play. After five waves: the firefight reads (legible FX, plunging fire, suppression), it sounds (1:1 procedural audio, night-light), it plays true (a living valley, a COIN pillar that discriminates patience from body-count), and it scales true — a squad reads as nine spaced men, not a 21 m giant blob. On the dead axes the change is qualitative, not incremental — which is what "5×" is supposed to mean. And every claim here is tied to a number from a headless harness or a feature you can see in a screenshot. Did an AI really build this? — the evidence, not the assertion, is meant to answer that.

The 5× Campaign Making a defense-grade war-sim READ, SOUND, and PLAY true

Headline scorecard — before → after

01The Method — orchestration as the differentiator

1 · Recon fan-out

2 · DO-NOT-RETRY ledger

The recon → wave → verify pipeline

File-partition map — why the waves could not collide

02Wave A — Combat Soul · the firefight now reads true

Hypothesis — the exchange ratio was inverted

Method — the headless harness

Change — six engine files, the right way

Result — before → after (with held-out)

Realism wins (realism-probe)

03Wave B — Procedural Audio · the absent transformative dimension

Hypothesis — the game was silent

Method — a 1:1 cue oracle

Change — a new lib/audio/ package on the render seam

Result — verified invariants

04Wave C — Sensory & Game-Feel · the valley sleeps, wakes, and weathers

Hypothesis — the valley was diurnally flat and weatherless on screen

Method — the atmospherics probe with a mover-faithful oracle

Change — render-side atmospherics + civilian brain

Result — headless-verified (deterministic)

Code-verified, capture in the real app

05Wave D — COIN Soul · "counterinsurgency is the real game" made true

Hypothesis — the strategic pillar was mechanically dead

Method — the campaign-loop harness, two scripted policies

Change — secure-build, two-way CERP, live directives, asymmetric promises

Result — the flip (default 3×8 tour)

The command layer — coin-hud.png

06Wave E — World-Map Scale · the squad stops being a blob

Hypothesis — the one clearly-visible realism gap left on the map

Method — a zero-dep scale probe with a true-footprint oracle

Change — figures track their footprint, squads aggregate into NATO icons

Result — before → after (scale-probe, seed korengal)

07The Charts — every headline number, drawn

08Annotated Screenshots — what you are looking at

09Honest Residuals — measured, characterized, logged

Wave A — suppression element-average peak below the guessed band

Wave B — sound quality needs a human in a browser

Wave C — weather is render-only; gameplay consequence deferred

Wave D — held-out 4×14 bimodal spread −3.0 (a harness artifact, not the engine)

Wave E — world-map scale: Bucket 1 RESOLVED; Bucket 2 + R3 deferred (issue 014)

10The 5× Assessment — how much better, axis by axis

Realism

Joy & awe

Change — a new `lib/audio/` package on the render seam