People Immersion · 2026-06-10 · Sim AI + Render + COIN

Making the people real

The owner wrote a standing brief — not a ticket, a dare: make every soldier, fighter, and villager move, behave, and interact so convincingly that a skeptical infantry veteran forgets he's watching a simulation. This campaign answers it with one architecture applied four times — group minds decide, individuals execute, everything is visible, and the valley remembers — and proves every piece with five new purpose-built probes, a live-app capture session, and a held-out seed set the work never tuned on.

The recon that corrected the brief

The brief's map of the floor (§3) was honest but partly wrong — and the campaign's first move was six recon agents sweeping the actual code before anyone designed anything. Three of the brief's claims did not survive contact:

• "Civilians respawn each day, identical, learning nothing" — false. They never respawned; they had been persistent named individuals all along. The memory infrastructure the brief asked for already had its subjects.
• "No cell leader, no coordination" — half false. Enemy cells already existed (squadId, casualty-shock contagion, leader succession) — what was missing was only the group decision layer on top.
• "Men don't react individually to a near-miss" — false. The suppression field is real physics; rounds crack past and pin men near the trajectory.

The genuine gaps were three: no group minds (every fighter an independent turret, every drill a metronome), no scenes (the shura a pop-up, the elder a name that never appears), and nothing visible (cohesion and brotherhood asserted in the data, absent from the eye). Those three gaps shaped the whole campaign.

Wave 1 — The enemy gets a brain sim · cell-combat.ts

Before: every fighter ran the same state machine alone — disciplined, but a row of independent turrets. Each man triggered his own ambush when his picture looked right, so a five-man trap opened raggedly over ten-plus seconds. After: every spawned cell gets a flagged leader, and cellFight makes the group decisions — the trap springs as one volley on the base gun's kill zone (FM 3-21.8's rule: the ambush is initiated by the most casualty-producing weapon — the per-man triggers defer to the leader), targets are distributed across the L from each man's own sight picture, fire-and-movement displaces by halves on the leader's clock, and a breaking cell peels to a shared rally, farthest man first, instead of scattering.

The proof is an A/B with a kill-switch: ITM_NOCELL=1 turns the coordinator off and re-runs the same seeds, so the delta is attributable to nothing else.

Cell coordination (6 seeds × 25 min, A/B)	Coordinator OFF	Coordinator ON
Ambush volley spread, p50	0.1 s	0.1 s
Ambush volley spread, p90 — the ragged tail	11.2 s	0.6 s
Breaking cell, mean dist to rally 30 s → 90 s	40 → 61 m (rout disperses)	54 → 34 m (peel converges)
Enemy rounds/min (gate ±15%)	baseline	−6.6%
Civilian melt tell (pre-contact vocabulary unchanged)	PASS	PASS — preserved by construction

The convergence row is the one to teach: you cannot fake it with a screenshot. A coordinated peel shrinks its spread around the rally point between 30 and 90 seconds after the break; an uncoordinated rout grows it. The probe measures exactly that second derivative of group behavior — the same way you'd tell a withdrawal from a panic on a real hillside.

The drill stops being a metronome sim · squad-combat.ts + friendly.ts

The brief's complaint: two squads assaulting the same objective moved identically. A new probe (drill-timeline-probe.ts) fingerprinted the rigidity first — bound step-off spread was literally 0.00 s σ: every buddy pair stepped off on the same tick, every time. Three mechanisms loosened it without touching the load-bearing clocks:

• Nerve — a per-man scalar from static traits only (composure-max, experience, aggression, hash-jittered, zero-mean by construction) that drives bound step-off hesitation, pin-threshold jitter, and cover-acceptance taste. Step-off spread: 0.00 → 0.10 s σ; two same-SOP squads' first assault commits now drift 0.4 → 6.0 s apart on the same objective.
• Pinned-revert — an assault whose maneuver element is majority-pinned >4 s falls back to suppress, re-develops (12 s floor), and re-commits preferring the other flank. The "we're pinned, go to ground" beat the brief said didn't exist.
• The securing buddy — when the medic arrives at a casualty, the aid buddy becomes the third figure in the scene: he kneels 2.5 m off on the threat side facing out, because in TCCC the first treatment is security, not the bandage.

The brotherhood becomes audible sim + render · the callout bus

Cohesion existed as numbers; the eye saw none of it. The fix is a diegetic callout bus: the engine emits short shouts at the moments that matter — "contact front!", "man down!", "covering!"/"moving!" on bound swaps, "doc — over here", "on me" at succession, "head count!" at the consolidate — through a ring buffer on CombatSim with per-squad dedup windows. Phrase variety comes from string hashing, never from the sim's rng, so two same-seed runs produce fingerprint-identical callout streams: determinism is asserted, not hoped. A React-free presenter (lib/render/callouts.ts, cloned from the audio mapper's warp-safe high-water-walk pattern) draws at most 3 small plates above the figures at tactical zoom.

Callout probe (5 seeds × 2 runs × 25 min)	Result
Two same-seed runs, full stream fingerprint	identical
Man-down coverage (witness-eligible casualties announced)	100% — 80/80
Dedup violations / spam	0 · per-squad ≤ 8/min
Held-out seeds (hold-0..2)	22/22 coverage, det OK, dup 0

Wave 2 — The seams of march and fight sim · tasks.ts + formation.ts

The brief singled out the transitions — march → contact → lull → resume — as "where realism lives or dies." The design survey agreed so hard that three independent proposers invented consolidate & reorganize separately; that convergence was treated as evidence it was load-bearing, and the three inventions were merged into one mechanism at the contact→lull seam (Law 6: one robust mechanism, not a pile of patches).

After a real engagement ends on the march, the squad now does what FM 3-21.8 says a squad does: collapses into a kneeling security ring, the squad leader physically walks the line, ammo is cross-levelled to the automatic weapons (the ACE report's real consequence), casualties are carried to the ring, and the head count goes out on the net — 30–90 s keyed to casualties and suppression, preempted instantly by renewed contact.

Transition + texture gates	Result
Consolidate follows eligible post-contact lulls (transitions-probe.ts)	60/60 · resume 100% · stuck 0
ACE ammo cross-level to auto weapons	fires when rifle>gun imbalance crosses threshold
Medic-scene third figure (aid buddy kneels facing OUT)	covered — after a dead-gate bug was caught, see below
March texture: point-man caution, per-man line meander, night/weather interval	paceEff 0.58 → 0.58 (unchanged) · blocked 0% · jitter 1°/s

The movement texture is deliberately subtle: the point man eases pace into corridor chokes and holds one beat at the mouth; each man wanders his own hash-keyed ±fraction of a metre off the wake line so the file stops being beads on a string; rain and darkness close the interval. The gate that mattered was the non-regression: all of it ships with the cohesion suite's pace efficiency byte-for-byte at the baseline mean and zero blocked legs — texture that costs nothing.

Wave 3 — The village remembers sim · the elder, the ledger, the children

The brief called this "the hollow heart": the elder had a procedurally generated name but never appeared; a civilian death was an attitude penalty, not a person. Wave 3 makes the village's people first-class agents and its memory persistent state:

• The shura is staged, not displayed. Every village's elder is a guaranteed real unit (first villager spawned; VillageState.elder carries his actual name; succession repairs the binding when he dies). A KLE physically summons him: he walks out with dignity, sits down 2 m from the squad leader's ring slot — and only then (t.elderMet) do the attitude drip, the elder's ask, and the decision modals flow. A 15-minute no-show proceeds via "the elder sends his regrets" — the absence itself a read. Gunfire always wins: a staged threat aborts the summons, latched (verified — the elder turned home, then fled the actual firefight).
• The blood debt has a name. Civilians belong to 2–4-person households. A civilian casualty of US fire writes a named grievance to the village ledger; at the next first light the household and the elder gather at the compound for twenty minutes and bury him — one log line, weight not spectacle. Unresolved entries floor effective sympathy (Pashtunwali's badal, capped); the dwell event becomes "The Blood Debt", by name; paying solatia settles the ledger by name. Grieving households clear the road from troops, always; their children never trail.
• The children are the meter. In a friendly village, curious kids trail a weapons-cold patrol at ~9 m standoff; in a hostile or grieving one they are simply not out near your men — the melt's quieter sibling, running on attitude. FM 3-24 and CALL's atmospherics handbooks treat children-near-patrols as the canonical positive indicator; here it is emergent, not scripted.

"Abdul Khan of Kandlay was killed by our fire. His household will remember."— the ledger entry, written by the engine at the moment of the casualty (lib/sim/world/world.ts · applyCivcasBacklash)

Village-wave gates (encounter-probe.ts + suite)	Result
Elder departs ≤300 s · sits down	+42 s of on-station · 4.1 m from the SL (live match)
No dwell event before elderMet · staged ambush aborts the summons	held · melt wins, latched
Funeral: household at the grave at first light · save/load mid-gathering	named entry · 3/3 kin resume the gathering
Kids trail-minutes, +40 vs −40 attitude (tuned / held-out)	2.82 vs 0.07 · 9.59 vs 0.00
Dwell bands untouched (KLE 1.63 h in 1.0–2.0) · diurnal + melt mechanism · civCas	verbatim PASS · PASS · 0

The probes did the real work — six bugs that never reached the player

This is the section a skeptic should read twice. The campaign built five new deterministic probes — cell-coordination A/B (with the ITM_NOCELL kill-switch), callout determinism/spam, drill-timeline fingerprints, transitions, and the encounter probe (shura/funeral/kids) — and they were not decoration: every one of them caught a real shipping bug before it shipped. A feature you can't measure is a feature you can't trust; here is what measurement bought, verbatim:

#	The bug the probe caught	The number that told the story	The fix
1	The doc-shout metronome. "doc — over here" re-emitted every tick a casualty lay untreated — a spam wall, not a brotherhood.	390 emissions in one run → 6	Latched per casualty, not per tick.
2	The half-volley. The displace clock started at spawn and expired instantly — half the cell went scooting before their first shot, gutting the disciplined initiation the wave existed to build.	volleyP90 18 s — inverted, worse than no coordinator	Displace clock arms only after a man's first shot.
3	The quantized hesitation. The per-man step-off delay was gated behind a slow reconsider loop, which quantized it straight back to zero — the feature measured as not existing.	step-off spread 0.00 s σ again	Moved into the per-tick execution layer (friendlyBrain), re-measured: 0.10 s σ
4	The dead medic-scene gate. The securing-buddy logic checked a targetId field that no writer ever set — syntactically fine, semantically a gate that could never open.	scene coverage 0 / 1048 treat ticks	Re-keyed to the proximity-matched aiding state.
5	The meet-geometry bug. The elder sat down at the shura while the log said "sends his regrets" — the summons walked him to a meet point 9.2 m from the SL's actual ring slot, outside the 6 m met-gate.	9.2 m vs a 6 m gate	Meet point anchored to the SL's real slot; live: 4.1 m
6	The medic who fired on his own patient. The first fix for #4 wrote targetId — and silently poisoned updateFiring's trust in that field: a weapons-free medic acquired his own casualty. Caught by a balance bisect, not a glance: KIA rose with the coordinator ON and OFF both, and the KIA-up/WIA-down signature (wounded men dying) pointed at the aid station.	KIA 3.75 / 3.00 (ON/OFF) → fixed → 1.25	The targetId channel reverted; scene re-keyed without touching the firing path.

Proven on seeds the work never tuned on law 3

Everything above was tuned on the survey-0..39 convention and dedicated probe seeds. The final proof ran on fresh, untouched inputs:

Held-out check	Result	Baseline / gate
balance.ts — 16 deployments × 50 min, hold-* seeds	KIA 0.94 · WIA 6.44 · enemy 4.38 · civCas 0 · 0 stranded	HEAD KIA 1.08 · WIA 8.58 — friendly KIA below baseline with a smarter enemy
cohesion.ts — survey-40..45	paceEff 0.57 · blocked 0% · jitter ≤2°/s	HEAD 0.58 — march texture costs nothing
Civilian melt — survey-44 / survey-45	OK / OK (children melt first; 0 shots in window; det hash identical)	the gold-standard tell, untouched
Callout bus — hold-0..2, 2 runs each	22/22 coverage · dup 0 · fingerprints identical	det + spam gates
Kids meter — held-out seed	9.59 vs 0.00 trail-minutes	tuned-set 2.82 vs 0.07 — the signal grew off-tune

Grounded where a veteran would check

• FM 3-21.8 — battle drills and consolidate & reorganize (the ring, the line walk, ACE, the head count); the ambush initiated by the most casualty-producing weapon (the cell's base gun opens, riflemen defer).
• TCCC — security is the first treatment: the aid buddy's job is the threat side, facing out, while the doc works.
• FM 3-24 + CALL atmospherics — children near a patrol as the canonical positive indicator; their absence as the warning. The melt (already shipped) and the kids meter (new) are the same grammar at two volumes.
• Pashtunwali — badal — the blood debt is named, persistent, heritable by the household, and settleable by solatia: exactly the ledger the engine now keeps.
• The lived texture — Junger's War and Restrepo: the shura's slow dignity, the head count after contact, the radio's flat grammar. Weight, not spectacle — the funeral is one log line on purpose.

How every claim was checked

Claim	How it was proven	Number
The trap springs as one	cell-coordination-probe.ts A/B via ITM_NOCELL, 6 seeds × 25 min	p90 11.2 s → 0.6 s
A break is a peel, not a rout	same probe — rally-distance convergence 30→90 s	40→61 m vs 54→34 m
Men hesitate like men	drill-timeline-probe.ts fingerprints, same-SOP twin squads	0.00 σ / 0.4 s drift → 0.10 σ / 6.0 s
The brotherhood is audible & deterministic	callout-probe.ts — 2-run fingerprints, coverage, spam	identical · 80/80 · dup 0
The squad consolidates and resumes	transitions-probe.ts	60/60 · resume 100% · stuck 0
The shura is real and gating	encounter-probe.ts shura + live capture, state-read	+42 s · 4.1 m · elderMet gate held
The valley remembers across save/load	mid-gathering serialize → loadWorld	3/3 kin resume the funeral
Nothing the player had got worse	balance/cohesion/melt/dwell/autonomy + smoke/tsc/build, every wave; held-out tail at the end	KIA 0.94 vs 1.08 · civCas 0 · 0 stranded · dwell bands verbatim

Engineering record: docs/progress/2026-06-10-people-immersion/ — the brief's recon-corrected CONTEXT, the synthesized plan with its adversarial-review amendments, verbatim HEAD baselines, per-wave evidence (after-w1/, after-w3/), the held-out corpus (holdout/), and the live-capture session with its staging disclosed line-by-line (after-w3/live/captions.md). Probes: scripts/cell-coordination-probe.ts, callout-probe.ts, drill-timeline-probe.ts, transitions-probe.ts, encounter-probe.ts. Touched: lib/sim/ai/cell-combat.ts (new — the enemy group mind), squad-combat.ts + friendly.ts (pinned-revert, nerve, the securing buddy), lib/sim/combat.ts (the callout bus), lib/render/callouts.ts (new — the presenter), lib/sim/world/tasks.ts (consolidate, the staged shura), formation.ts (march texture), create.ts (the elder, households), world.ts (funerals, succession repair, the reception push), civilian.ts (summons, kids), events.ts (The Blood Debt), director.ts (leader flagging). Brief: docs/briefs/movement-and-interaction-immersion.md. Deferred: docs/issues/026. Every number quoted as measured; the residuals are in the red box, not a footnote.