HOW IN THE MOUNTAINS WAS BUILT

The squad declared "on station" 300 m from the village

A player ordered a squad to the village just right of the gate. It hugged the perimeter wire, bumped it, and "accomplished its mission" without ever reaching the village. On the default seed korengal it stalled 152 m short and called itself on-station 300 m from the objective.

The shape: metricize first — a headless harness reproduced the failure across seven adversarial seeds, four root causes were confirmed against the live code with dumped data, then all four were replaced by one unified "navigate globally, steer locally" locomotion layer.

"Navigate globally, steer locally" — one locomotion rebuild to end four layered bugs hiding in plain sight.

The navigator dutifully walked a 718 m spiral to a 391 m objective

Squads sent to a far village would orbit the outpost in a spiral, string into a 100 m accordion snake, and spin in place at over 400°/s at the first awkward waypoint. The pathfinder was returning routes that looped 1.5 times around the COP to connect two points 15 m apart.

The shape: two new harnesses turned four bugs into hard numbers, then a corridor-constrained full-resolution A* replaced the broken coarse-plus-repair pathfinder, and a wake-following cohesion model replaced the flickering breadcrumb follower. Verified with before/after trajectory renders.

The navigator dutifully walked every metre of a 718 m spiral to reach a village 391 m away.

Engineering record: docs/progress/2026-06-04-movement/.

64 candidates, 10 survivors: waking the dead code

The sim was honest in skeleton but full of finished systems nobody had switched on — enemy indirect fire, IED teams, and spotters built with zero callers; civilian casualties tallied but never read; wind stored but touching no ballistics. A 5.56 round wounded the same at 800 m as point-blank, and the insurgency could be ground to zero by attrition alone.

The shape: fan-out discovery — 10 subsystem-expert agents proposed gaps, a judge ranked 64 candidates to 10 (rejecting anything already modelled), each measured with a probe before/after, then hardened by two adversarial review waves that ran their own experiments.

Realism by mechanism, not by fiat — the right casualty ratio fell out of the systems interacting.

Engineering record: docs/progress/2026-06-04-realism/.

Blue rectangles to bas-relief: 160 hand-authored SVGs remake the map

The tactical map was primitive canvas shapes — blue rectangles for soldiers, wireframe boxes for buildings, coloured dots for everything else. Nothing about the visual layer said this was a defence-grade infantry sim. The outpost looked like labelled wireframes; villages were a house glyph and a name.

The shape: a background multi-agent workflow with one agent per sprite family — each rendered its own SVG and critiqued the rasterised PNG before returning — followed by a four-lens art-director pass driving a fix round. Verified live across every zoom band.

Every asset was seen and critiqued by the agent that drew it before a human ever looked at the screen.

The router was fine — fatigue was eating the march

Squads were "setting up short" 273 m from their objective, or stalling halfway across the valley. Only 36% of physically reachable villages were ever reached, and just 13% of those directly opposite the gate. Everyone assumed the A* router was broken.

The shape: a fair oracle — an adversarial harness scored real-sim arrival against an 8-connected BFS ground truth, and per-tick traces apportioned blame to the exact tick. That overturned the hypothesis: fatigue saturation, not the router, was eating the march.

The router was already good — a router rewrite would have fixed nothing.

A frozen turret where a security halt should be

Squads setting up 360° security marched men straight through each other to reach raw ring slots — and once halted, soldiers snapped their facing at 700°/s and froze, a dead turret instead of a scanning perimeter. Spacing was double doctrine. The whole thing read as mechanical, not military.

The shape: metricize first with two harnesses, then a three-expert research workflow (FM/ATP 3-21.8, game-AI steering, crowd-sim) converged on the fix — slot assignment by nearest, a settle that scans rather than snaps — re-measured in isolation, determinism confirmed.

"The hardest part of command is watching" — the 360° now forms like a real squad, not a lattice snapping into place.

Engineering record: docs/progress/2026-06-06-movement-realism/.

The squad stopped dropping mortars on itself

A player report: "the squad called a mortar mission much too close to them, nowhere near the enemy." Both halves were real — the AI averaged all enemy positions (landing rounds on the squad in an L-shaped ambush), and the danger-close gate existed but never actually stopped the call, producing a genuine US fratricide.

The shape: metricize first — a probe intercepted every AI call-for-fire across 40 deployments, scoring the aimpoint against independent ground truth, then let rounds fall to count real fratricide. The fix: a densest-cluster aimpoint, a real danger-close withhold, and an FDC check-fire.

It found a one-line averaging mistake and a missing safety the way a soldier would: by measuring where the rounds actually fell.

The strategy layer that punished patience and rewarded body-count

The game's central claim — "you can win every firefight and still lose the valley" — was mechanically false. Careful counterinsurgency and a pure body-count run scored almost identically. Projects never completed because no element stayed to hold security, so every funded build sabotaged itself. CERP only counted toward zero. Directive deadline-and-penalty code was never read anywhere.

The shape: a campaign-loop harness compared careful-COIN vs body-count across seeds and game-day windows as the authoritative ledger, then a coherent engine wave fixed each root cause in isolation — and a later pass fixed the instrument itself, which had been over-marching careful patrols into ambushes.

The harness meant to prove COIN worked was itself broken — it over-marched careful patrols into ambushes, doubling KIA and inverting the spread.

Resolved tracked issue 015. Engineering record: docs/progress/2026-06-06-coin-real-game/.

The silence that made you realise: no game sounds here at all

The game had no audio whatsoever — no crack-thump of incoming rounds, no SAW rip against a PKM hammer from the high ground, no mortar shot-and-splash, no radio net under the firefight. The genre's spine was missing, and the player was watching a silent simulation.

The shape: ground-up procedural synthesis with no binary assets and no npm dependency — a strict pure/browser split (a deterministic cue collector in the sim half; Web Audio node graphs in the browser half), verified headless by byte-identical replay and live against a debug ring buffer.

Every sound is synthesized procedurally — no binary assets, no npm audio dependency.

Engineering record: docs/progress/2026-06-06-procedural-audio/.

Five waves, one campaign: the firefight, the audio, the valley, and the soul of COIN — at once

The game still played like a prototype: US soldiers died at a 5:1 disadvantage, the strategic layer was indistinguishable from a body-count run, soldiers were painted 36× too large, and the valley was silent and diurnally flat. This pass tied the loose threads of the day into one verified campaign-wide release.

The shape: a goal-driven orchestration — a 12-domain recon fan-out produced 63 ranked hypotheses, then five file-disjoint implementation waves (combat soul, audio, game-feel, COIN, scale), each ending in its own adversarial verify gated by the full done-gate suite.

You can win every firefight and still lose the valley — and now the scoreboard finally proves it.

From mono-and-silent to a living canyon: a 20× soundscape rebuild

The earlier audio pass put sound in the game — but it was combat-only, essentially mono, had no reverb, and fell silent between firefights. The valley didn't breathe. This pass rebuilt it into a living place: a canyon that echoes, a wind/river/generator/wildlife bed keyed to the sim's time-of-day and weather, terrain that muffles fire behind a ridge, and 5-layer weapon synthesis with a correct 5.56-vs-7.62 timbre tell.

The shape: a 6-agent research fan-out (gunfire DSP, canyon reverb, ambient design, spatialization, adaptive mixing, valley acoustics) → an additive bus architecture → three parallel module builds → an offline oracle (audio-render.ts) + spectrogram viz tuned to all-green. The adversarial pass caught a real oracle bug (a voice counter that never decremented, so the firefight was measured on only its first 32 cues).

100% procedural — Web Audio only, zero audio assets, the reverb impulse is generated noise — and deterministic.

The command surface, re-cut: 461 → 1 measured UX defects

The HUD a commander stares at for a whole deployment was dense and handsome — and quietly hostile: built on 9–11px text, with zero keyboard-focus visibility, no motion-safety, glyph buttons a screen reader calls "button," and the campaign's five north-star metrics drawn as identical 2.5px slivers. This pass kept every ounce of the milspec soul and fixed the legibility and accessibility under it.

The shape: a CDP audit harness metricized the debt at 461 defects before any edit; an 11-agent survey (5-persona judge panel + 6 single-axis proposers + a synthesis lead) produced a deconflicted backlog; four atomic rounds implemented it; a 5-agent adversarial paired re-judge graded the result and demanded residuals. A focus-floor + a 12px type floor did most of the collapse; an overflow oracle and a keyboard-behaviour harness (6/6 pass) verified nothing broke.

The literal 20× lives on the objective axis — a 461× defect reduction, reproduced deterministically — and an independent expert panel unanimously confirmed the direction and magnitude.

The combat outpost, made real: a fire plan sited by terrain, a garrison that lives, and the watch

The COP was geometrically correct and dramatically inert — a cluster of b-huts on a tan patch, defended by eight identical nubs, with crew-served guns handed out by array index and no Mk19 at all, a garrison that loitered, and a wire attack you could only watch as a cutscene. copaudit was already 100% green, so this was never a bug hunt: it was taking a system that works and making it convincing.

The shape: a six-specialist design survey (defense, visual, garrison, facilities, gameplay, atmosphere — each pinned to ATP 3-21.8 and the real file:line facts) and a synthesis judge produced a Law-6 unification: one terrain LOS-sweep spine that sites the M2 on the longest avenue and the Mk19 over the worst dead ground, feeds the render silhouette and the mortar pit; one render env seam for the night life-signs; and one reused fire-request loop that turns a wire assault into a Final Protective Fire the commander must clear. Seven atomic commits, each verified.

The art to build a recognisable COP already sat unused in the sprite manifest; the doctrine to defend it sat in the terrain. The pass wired the engine to the ground truth — and copaudit stayed clean (0 sealed pockets ×9).

Reading the ground: traversable steep terrain, footpaths that mold, cliffs you can read, cover you can use

The map looked alive but quietly lied about itself. A new probe floods the valley the way the mover does and found the tell: 80% of the ground was passable but only 48% was reachable — the hard "no foot traffic above 51°" cutoff salt-and-peppered fake cliffs across climbable faces, shattering the walkable map into islands. The footpaths were a 5 m paint smudge; a 56° wall was drawn the same warm tan as a meadow; and the cover that decides a firefight was one averaged number per cell, decoupled from the rocks you could see.

The shape: four asks, four measured iterations, each baselined and re-measured in isolation. Soften the impassable line to reconnect the steep band (reach 48→61%, real cliffs preserved). Switchback foot-trails up the spurs (a benched first version cut impassable trenches — caught in a trail render, rebuilt to surface-lay). Stroke every captured path centerline as a scaled dirt line, and shade everything above the foot-passable slope as sheer cool rock. And promote the boulders from cosmetic sprites to sim objects the renderer draws from — so the rock you see on rocky ground is where the engine's cover is. (The combat-cover half on open ground was measured and cut: even a light stamp dragged firefights +89% WIA — deferred to a sub-cell model.)

The restraint was the work. A cost-floor tweak that felt right bought no routing win, ran the sim 1.5× slower and spiked WIA 71% — reverted on the numbers. A cover change that looked fine in aggregate turned ridge firefights into grinds — caught only because a same-seed A/B disagreed with the cross-config one. Cross-config comparisons lied twice; same-seed A/Bs told the truth.

Climbing the mountain: a squad that switchbacks up a face instead of ringing the spur

Order a squad to a peak OP and it walked a 3-kilometre arc around the mountain instead of climbing it. The cause was geometric: the foot-movement cost was isotropic — it read a cell's slope magnitude, the same in every heading — so a diagonal traverse bought nothing and a switchback could never be cheaper. And the planner was 8-direction, so the shallow traverse angle a switchback needs fell between the grid headings. Two earlier in-place attempts had been reverted for regressing every route and stalling a squad.

The shape: the scoped rebuild the issue had already named, built to be additive. The slope penalty moves out of the per-cell cost and into the edge — speed now depends on the signed grade along travel, so a long switchback genuinely beats a short scramble (it minimises travel time, not length). A Theta* (any-angle) search lets a leg run at any heading — the geometry the 8-dir grid couldn't draw — with a turn penalty for few clean bends. It fires only for a deliberate squad march to a steep, elevated, tactical objective; world generation and valley-floor routing never trip it, so the terrain and village movement are byte-identical. A same-seed A/B kill-switch proved it: planner off = byte-identical to HEAD.

The honesty is in the residual. OPs the probe tucks behind a genuinely impassable massif still detour — and that's real terrain, not a planning miss, so the planner correctly falls through to the proven router rather than paying for a near-whole-map search to shave one adversarial seed. And the balance cost is stated plainly: climbing exposed high ground costs a little more (KIA 0.92→1.08) — the realistic price of the climb, accepted rather than dodged, with the stall guard the prior attempts failed now passing.

Behind that rock: directional, posture-aware cover a soldier can actually use

The drawn boulders became real sim objects in June, but the firefight still read a single averaged cover number per 5-metre cell — so a man tucked behind a rock was scored exactly as exposed as one standing in the open (measured: behind 0.08 = open = flank). The rock was decoration to the bullets. And the obvious fix had already been tried and reverted: adding open-ground cover omnidirectionally made both sides survive longer and dragged every firefight into a +89%-WIA grind.

The shape: the missing thing wasn't more cover — it was direction. A boulder only stops the round coming from the bearing it faces. So a discrete object between shooter and target now blocks a posture-scaled fraction of the shot (a low rock hides a prone man far more than a standing one), on the fire-hit path only, and the AI of both sides moves to tuck behind it. Because a flanker's round is never on that line, the squad's already-shipped autonomous flank defeats the cover, and the fight resolves by maneuver instead of a longer attrition clock.

The reverted attempt added cover and casualties rose; this one adds directional cover and casualties fall — US WIA −22% on the tuned seeds, −36% held-out, with the enemy still dying via the flank. Same idea (open-ground cover), opposite result, because of one property: a boulder faces a direction.

Supplies that bite — and the COP assault that wasn't there

Issue 021 wanted the COP's fortification (fob.hesco) to matter in a fight: a hardened wire should blunt a costly assault. Metricizing it first — Law 1 — flipped the whole campaign. A probe staged the complex attack and measured garrison casualties hardened vs neglected: zero either way, across 8 seeds and a 20-minute trace. The trace showed why — the attackers spawn at 340 m and never close the wire. The insurgent AI engages and shoot-and-scoots laterally; there is no assault behaviour at all. A "complex attack" is a standoff from the ridges (the real Korengal), so fortification and claymores have no event to defend against.

The shape: rather than fake a coupling to a non-event, ship the part of 021 that affects the gameplay you actually have — the patrol. Water, food, batteries, and medical drained every day to a floor with zero consequence. Now each bites, as a bounded clamp: dead batteries cost the US their night-vision edge (a 2.4× detection penalty in the dark), dehydration and underfeeding halve fatigue recovery, and a bare aid bag slows wound recovery — while a fully stocked COP plays exactly as before.

The most valuable line of code this campaign was a probe, not a feature. It proved the assault the fortification was meant to stop never happens — saving a coupling to a number that never matters, and pointing at the realism win that does: the resupply you forgot to call.

Reading the sun: vegetation that knows which way a slope faces

The valley grew the same vegetation on a north slope and a south slope — no notion of aspect, the single most visible fact of a real mountainside (shaded slopes hold moisture and grow forest; sun-facing ones bake to scrub). An aspect term in classifyLand fixed it — and because vegetation is cover and concealment, it could not be slipped in. A three-strength balance sweep showed the firefight moves chaotically with the terrain (KIA non-monotonic, strandings flickering), so it ships at the one strength that earns its place.

The shape: measure, don't assume. 0.16 ran too hot (+56% KIA, stranded); 0.025 was chaotic (also stranded); 0.05 clears the no-stall guard, gated to the steep faces so village/COP siting stays byte-identical. Held out on fresh seeds: KIA down on both sets (the permanent loss never regresses), the lone stranding pre-existing (aspect adds none), and the wounded count terrain-dependent — disclosed, not buried.

It shipped on by default only after the balance earned it — three strengths measured, a held-out check, and the honest caveat (combat intensity varies with the ground, which is itself realistic) stated in the open. The realism is in the vegetation; the rigor is in refusing to ship it until the numbers said it was safe.

Closing the open issues — four realism wins shipped, two reasoned closeouts, zero regressions

A single pass over every open issue, run the way this repo demands: metricize before touching code, prove every win on held-out seeds, adversarially verify before "done" — and ship only what the numbers earn. Four realism wins shipped active (019 switchback pathing, 020 directional cover, 021 logistics teeth, 007 aspect-driven vegetation), each with a held-out proof and — for the two combat changes — an adversarial workflow that tried to break it and couldn't. The fourth was the hard one: aspect touches the cover/conceal field a firefight reads, so it took a three-strength balance sweep + a held-out A/B to find the one strength (0.05) that improves the sim — forest-faces-north 59%→64%, scrub-faces-south 48%→56%, KIA down on both seed sets, no aspect-caused strandings, siting byte-identical. It ships on by default. The two remaining tickets (011, 009) already run their active fix in the shipped game — a progressive, cached relief bake behind the loading screen; the reachability-aware spawn-snap + benching guard — so what's left is a logged do-not (resolution-lowering would degrade the relief the realism work added) and a documented terrain floor, not unbuilt work.

The headline is that six issues now carry a live fix and the sim got better without a single shipped regression — including the one (aspect) that took a sweep across strengths to ship honestly instead of guessing. On a project whose success metric is a skeptic asking "did an AI really build this?", the discipline to keep sweeping until the numbers point to the strength that ships — and to hold a perf ticket at its honest deployed mitigation rather than fake a further win — is the part that's hardest to fake.

The mixer and the whistle: per-category sound control + five audible upgrades

Two asks: let the player turn categories of sound on and off, and make the sound itself better. The mixer rode the existing bus architecture — one user gain per category (combat / ambience / radio / alerts) inserted between each bus and master, because the contact-duck system writes absolute values to the bus gains and a user trim on the same node would be stomped on every duck. The valley reverb return joined the combat category (muting combat must mute its echo), the danger-close klaxon moved to alerts where it semantically lives, and an oracle re-render proved the re-plumbing level-neutral: all 24 scenes within 0.15 dB.

The shape: the five audio upgrades each came from a recorded residual or a missing battlefield signature, and each closed against the offline render oracle. The big one was a synthesis subtlety: the calm bed measured near-mono (corr 0.997) despite panned voices, because every bed loops one shared noise buffer and two in-phase filters of the same source stay correlated no matter how they're panned — the fix was starting each voice at a different offset into the loop, and only then did the panning (river bands straddling the river's real bearing, the generator on the COP's, Haas rain) buy real width. The whistle is pure determinism: the sim already counts down a fire mission's etaS, so the mapper voices one whistle 2.4 s before the first round, once per mission, latched and warp-safe.

The honest residuals lead the report: the band-split river left the calm bed ~2.9 dB quieter (in band, knob named), and there is still no MEDEVAC helicopter — evacuation is instantaneous in the sim, so a rotor loop has nothing to attach to, and faking one with a timer in the audio layer was the change deliberately not made.

Calibre voices: every weapon gets its real sound

A soldier tells weapons apart by ear before he sees anything — the report's pitch and weight ARE the information. But the audio layer only knew faction and "is it a machine gun", so an M2 .50, an RPG launch, an M9 pistol and an M4 all collapsed into two rifle cracks. The unlock was one sentence of sim code: the engine now stamps which weapon produced each muzzle/blast effect, and the whole calibre ladder fell out of it — measured, solo/near: M9 7701 Hz > M4 6860 > M24 6450 > Enfield 5876 > DShK 5242. The .50s needed one more honesty: the shared 6 kHz attack transient had made every gun equally bright, so the attack itself is now weapon-tinted — the M2's is a low whump.

The shape: new sounds where the battlefield was silent (the RPG's pop-and-whoosh launch — "RPG!" is finally audible as a launch, not just an arrival; the M320's bloop and the Mk19's bolt-thunk; a man two metres away swapping mags at −51.6 dB), per-weapon voice rows refining the class voices (bolt guns cycle their bolts ~0.5 s after the report), and physics upgrades to the loudest moments: the IED now heaves (seismic 26→16 Hz ground wave leads, the airborne crack arrives 8 ms late through the soil cap), the incoming shell tears instead of singing (two incommensurate modulators that never phase-lock), the near miss completes its N-wave. Where the standard metrics went blind — centroid and HF% are bin-count-dominated by broadband noise — the oracle grew a low-band share metric, because a blast's calibre lives below 200 Hz.

The residuals are named: the RPG launch never fired organically in the three held-out runs (gunners carry one round) — routing proven by direct mapper check, organic occurrence unobserved; and there are still no synthesized screams, deliberately — the recorded negative on synthesized radio voice ("reads cheesy faster than squelch reads sparse") extends to wounded men, with higher stakes.

The valley is real — a WebGL terrain rebuild (attempt #2)

The first attempt at this overhaul was reverted: it moved the existing flat painted relief onto the GPU byte-for-byte and laid a colour filter and a moving sun over it — the same map with the sun switched on. This rebuild inverts the idea. The painted albedo is demoted from "the image" to one input of five, and a WebGL2 fragment shader recomposes the surface per-pixel from the simulation's own arrays — landcover, slope, cover, the heightfield — against a procedural material library, lit in linear radiance by the live master-clock sun.

The shape: a WebGL2 underlayer beneath a transparent Canvas-2D HUD (the units and contours are a separate, never-graded layer — the legibility firewall). Three passes a frame: terrain → RGBA16F HDR (per-landcover materials + detail-normal raking the live sun, baked horizon AO, multi-km ridge shadows, a dark-silt flow-advected river, aerial perspective + in-fog god-rays), then threshold bloom, then ACES tonemap + time-of-day grade + dither. A u_detailGain zoom-ramp collapses the whole transformation to byte-faithful relief at strategic zoom, so the band the old map got right is provably preserved while close zoom resolves into real ground. Nine commits, each green; reuses the verified Kunar solar model, the bit-faithful heightmap port, and the exponential cast-shadow march.

Residuals named, partial stated as partial: the terrain got the full 10×, but the assets got grounding (sun-tracked shadows), not a full relight — the 164 sprites are still flat bitmaps. A per-sprite normal/AO relight and FX particles are a scoped follow-on (issue 028), logged rather than shipped half-done. The win the verification earned: a four-critic panel caught a snow-white river and a contrast-crushing haze the build's own author had rationalized away — fixed, then re-confirmed by a fresh reviewer.

When "too squiggly" meant "too straight"

The owner named four realism complaints: soldiers get stuck, the paths are too squiggly, the terrain is too smooth, the outpost feels fake. Metricized first, two inverted. The "squiggly" paths were a near-beeline — planned route ratio 1.12, 66% of it straight down the fall line — and the visible wobble measured 0.16 px at play zoom; real trails needed to traverse, not de-wiggle. And "getting stuck" wasn't the patrols (router within 0.6–5% of optimal) — it was an in-combat freeze where a fighter breaking contact got a beeline into a cliff and re-issued the dead target forever: one unit frozen 678 s.

The shape: three build fronts, each grounded in cited doctrine — the uphill march re-anchored Naismith→FM 3-97.6, the foot-trails to the USFS half-rule, the walls to real Korengal gneiss-and-schist bedding (Junger, Restrepo). Then two commits fixed the test: a single unused RNG draw swung the COIN gate ±19 pts (relief-of-command is an opening-days lottery), so it now scores the paired best seed. Honest residual: the KOP was dressed, not reshaped — the perfect-circle generation rebuild is the next campaign.

The enemy gets a name

The valley could simulate every bullet, but there was no one to hunt: the insurgency was two scalars and a tempo clock. Fighters materialized, fought, vanished — an escaped fighter cost the enemy nothing — and the game's sourced, reliability-scored intel feed pointed at weather. Now 3–5 persistent cells live in the draws: named leaders who survive between fights (kill one and the cell renames, weakens by a third, and remembers), munitions caches that IED ambushes must drain, and a patrol-heat grid that learns your habits — predictable routes get IED'd. The population is the sensor: a won-over village names its cell, then locates it, then maps it — the Enemy Picture panel, the map markers and the new weekly commander's assessment all read one shared intel gate, so the player sees exactly what the fiction has earned and never ground truth.

The shape: same session, the other rage-quit died — relief-of-command now reads an attributed evidence file (every confidence dock tagged casualties / civcas / directives) and fires only on a nameable pattern past a five-day grace, never a day-3 dice roll. And the honest part: the build shipped a conservation bug — exfil deposited +1 into a roster the fighter never left, printing strength 64→80-cap in a single hot game-day, and the acceptance probe asserted the bug because the contract's wording was wrong. The skeptical pass caught it; exfil is net-zero now, KIA exactly −1, and the probe proves conservation both ways.

147 MB of transcripts distilled into one file that every agent reads first

After 23 real build sessions the project had accumulated 147 MB of transcripts encoding hard-won patterns — verify-as-a-number, fan-out orchestration, adversarial passes — but none of it was reachable to a fresh agent starting cold. Every new session re-discovered the codebase and sometimes re-proposed approaches already measured and rejected.

The shape: a five-workflow pipeline (Mine → Orient → Generate → Tournament → Finalize) extracted the corpus signal, generated five candidate operating philosophies in parallel, ran each through an adversarial critic and a 3-judge panel, then synthesized a champion from the best spine with grafts from the runners-up.

A CLAUDE.md is not documentation — it is latent-space conditioning: mined, distilled, and won.