The prop library was costing players more than it was worth. hunyuan3d
hands back ~120,000 triangles and 1024px textures per object, so a
COOKING POT was a 4MB download — and the fryer bench, which loads a pot,
a basket and a rack, made a phone fetch 13MB before it could show you
anything. Right after shipping touch controls, that is the wrong bill.
scripts/optimise-models.sh: simplify to 10% of the triangles with a
0.002 error bound, textures to 512. Across 51 models that is 214MB →
47MB, and at the distance this game frames anything, the before and
after screenshots are indistinguishable — I checked the pot at full
frame before trusting it on the rest. The fryer bench now costs under
3MB.
The script skips anything already under the threshold, so a second run
cannot grind an optimised mesh into mush; delete the file and re-run
gen-assets.sh to go back to an original.
Regression sweep after: toast scene pristine, grill 9.4, pan 8.0, steak
9.4, bruschetta 9.6, air fryer 9.5, chips 10.0, pasta 10.0, eggs 10.0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The egg bench gets its real ceramic mixing bowl. Everything else in this
pass is subtraction, because a prop that fights the game is worse than
no prop:
- The generated tumbler renders OPAQUE, and that glass exists for one
reason: watching an egg sink, stand or float in it. A pretty glass you
cannot see through deletes the float test. Reverted to the see-through
cylinder.
- Six models (wire basket, raw/cooked wing, raw/golden chip, egg carton)
were shipping in the bundle unused — 27MB of dead weight served to
every player for nothing. Deleted, and their lines in gen-assets.sh
are commented with WHY they were rejected so nobody regenerates them
by reflex: per-piece food must tint per frame (colour is the verdict),
and the top-down rack must never occlude the wings.
Verified: eggs still 10.0 with the new bowl, float test still legible.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The MODELBEAST station lane finished — 14 hero props, all local, all
free — and the single-instance vessels are now wired into the benches
that were still cooking in grey primitives:
- the deep fryer: a steel pot with handles, a wire basket that rides the
lift animation, and a proper drain rack stage left
- the poach: a perforated slotted spoon the egg actually sits IN
- pasta night: a real stockpot, with the salt visibly clouding the water
Per-piece food (wings, chips) stays procedural on purpose: their colour
IS the gameplay read — raw to golden to regret — and a textured mesh
tinted per-frame reads as mud. Fidelity loses to legibility.
Three fixes the screenshots forced, none of which guessing would have
found:
- The oil disc was sized to the understudy cylinder, so it poked out
past the real pot's narrower rim like a yellow skirt. Refitted.
- The basket sat below the oil line — invisible, and a basket you
cannot see is a basket you cannot judge. Raised proud, handle kept.
- The spoon is 1.9 long in Z with its bowl at one END, so centring the
model parked the bowl a full half-length away from the egg. Measured
the bounding box in-browser and shifted by half the length; the egg
rides the bowl now.
And one prop rejected: the generated egg carton ships with six eggs
baked into the mesh, and that bench is entirely about which eggs are
YOURS — draggable, float-tested, state-coloured. A carton that already
has eggs in it is a lie the player has to see past. An empty tray is
queued instead.
Verified after: chips 10.0, poach 9.9, pasta 10.0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Pasta night, and the kitchen's first genuinely irreversible verb. Salt is
a HELD pour, not a click: hold S and it falls, you stop it and nothing
else will, and `salt` is a number that only ever goes up. The words walk
you to the edge and then stop helping — 'shy of it, more while you still
can' becomes 'seasoned like the sea. stop.' becomes 'heavy handed. there
is no going back.' Salt thrown in after the pasta counts a quarter: it
seasons the water you are about to pour down the sink.
The rest is the atlas's pasta doctrine: pasta wants a ROLLING boil, and
a lazy simmer glues the strands into one brick because they sit in their
own starch instead of tumbling. Al dente is a window in immersion
seconds that never appears on screen. And the last verb is restraint —
RESERVE a cup of the starchy water before you drain, or the sauce has
nothing to hold onto.
Two things the adversarial pass caught, both about stakes being fake:
- Brine scored 8.6. An inedible plate cannot be averaged away by good
timing elsewhere, so ruin now falls off a cliff (not a ramp) and a
brined pot is CAPPED at 3.0 — the one verb you could not take back
gets the one score you cannot climb out of.
- The sourdough starter was buttering up chip day. Her 'is this a toast
day' test was a blocklist of stations that existed when she shipped,
so air fryer, fryer and pot all silently qualified as toast. Replaced
with stationOf()/isToastDay() in orders.ts — every station flag listed
in exactly one place, where the next station cannot slip past it.
Verified: honest pot 10.0; brine 3.0 at the cap; the lazy cook (simmer,
salt after the drop, drained every drop) 4.8 with all three rows naming
the exact sin.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Day 12: juice order routed end-to-end (72.8%, 4 pips, THE JUICE beside
THE TOAST on one scorecard). Day 13: the stop-line — proper dice stops
at 0.846 and scores 1.0; through-the-root scores 0.0 at identical CV.
Sting 0.098 quick → 1.0 after 10s idle; the screen blurs to genuinely
unreadable (screenshot-verified). Juicer GLBs landed mid-review.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Juicer verified through __tj: Pyramid 72.8%/4 pips, Juicernaut 100%/all
caught, damaged halve caps at theoretical 0.6, mash 8.5% + pulp,
reversals 77.1% vs 72.8% equal-sweep. Station renders (cone, moat,
filling glass, live HUD). Not yet routed into a day — integration point
documented in sim/juicing.ts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- button_mushroom / portobello / king_oyster in the ingredient cast:
slippery skins (0.55-0.8, the tomato treatment), soft flesh, portobello
gills layered and weeping; new 'mushroom' cap-and-stem silhouette.
- Verified via harness: dinner knife pressed on a button skates (1 slip,
cut lands 0.157 off-aim; tomato baseline 2 slips / 0.218), The Best
Thing bites clean through under the same press, sawing properly cuts
center with zero slips.
- gen-assets.sh: find MODELBEAST at ~/MODELBEAST, token from
backnforth/.env, MB_HOST to the m3ultra queue; every mb stage retries
10x with 60s pause because m4pro's venvs are gone (ponytail-marked,
remove when m4pro is healed).
- launch.json: port 5273 (5173 taken by another session's server).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A particulate spread. The jelly is ordinary rheology; the rind is discrete
solids that ride the knife and drop off per unit of stroke DISTANCE, not per
second — a dabbing knife makes a pile, a moving knife lays a trail, and since
the shed rate scales with what's left on the blade, one dip cannot cover the
slice. Even rind takes several dips placed with intent. Scraping picks bits
back up (60% survive the blade), so a clump can be argued with.
The judge scores the scatter with the Clark–Evans nearest-neighbour index over
the bread's own area, and quotes it: R near 0 is one clump, 1 is random, above
1 is deliberate. Verified end to end on day 8:
dab in a corner 8 bits "one clump (R 0.24)" drags grade to B
three placed dips 24 bits "evenly strewn (R 1.17)" A, 8.9/10
First cut shed per-second; even careful strokes dumped the whole load in the
first half-second and scored "clumped (R 0.50)". Distance-based shedding is
what makes evenness the reward for technique rather than luck.
Rind renders as instanced chunks parented to the slice mesh, so it rides the
pop, the landing, and the pedestal turn. Points, not a field: the judge wants
nearest-neighbour statistics, and those want points.
Same sRGB trap as M3, new spot: the instanced material's colour was passed as
raw floats, which the renderer reads as linear — candied orange rendered as
pale butter. setRGB(..., SRGBColorSpace).
New content: day 8 handwritten (Deidre wants a piece in every bite), marmalade
in the procedural pool, five rind-district lines for the judge, and a marmalade
jar GLB generated on-device like everything else, with a plain-glass fallback
while it loads.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Playable: pick a bread, set the dial, drop the lever, watch it brown, pop it
onto the plate. No timer readout — you go by smell.
- toasting.ts: heat map from real toaster behaviour (element stripes, cool top
edge where the slice stands proud of the slot, edge falloff, per-run noise
bias). Moisture must boil off before the crust browns at full rate, so
sourdough stalls then catches up; sugar browns and burns early.
Rate is measured against the heat map's actual ~0.77 mean, not guessed:
power 6 -> golden 0.511 at 10s, verified in-browser.
- Props on a real scale (1 unit ~ 11cm, one slice wide) — the toaster is
meant to dwarf the bread.
- The pop solves a genuine ballistic arc from slot to plate.
Three bugs found by driving it rather than trusting the compile:
- valueNoise2D read one past its grid at x=w-1, so heatBias was NaN along an
edge, which poisoned mean browning to NaN — i.e. every judge score would
have been NaN. Clamped.
- erode() clamped at the field border, so border texels could never erode out
of the mask and kept those NaNs in scope. Off-grid now counts as outside.
- Timer.connect(document) zeroes dt whenever document.hidden is true, which
froze the sim solid in an embedded browser. Dropped; rAF already pauses for
hidden tabs. App.step(dt) is now exposed so the game can be driven
deterministically for verification.
Also: screen shake was mutating the camera permanently instead of offsetting
it per-render, and the plate's lathe profile touched the axis (degenerate
triangles -> starburst normals) — rebuilt from primitives.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The bread slice is the whole game's canvas, so it's built first and properly:
- Field: a 128x128 scalar field with bilinear sample, a mass-conserving brush,
self-levelling relax, and the masked statistics the judge will read.
- Slice: extruded loaf silhouette (domed top), planar-projected UVs so the sim
grid lines up with the bread, per-texel silhouette mask, and four fields
(browning/dryness/spread/damage) packed into one RGBA8 texture per frame.
- Custom shader: browning ramp with blotchy char, crumb/crust/inclusions,
spread wetting + opacity + thickness relief, gloss, gouge exposure, and
flat data views for the judge screen.
- bread.ts / spreads.ts: the archetypes and the rheology the difficulty
curve will live in.
- scripts/gen-assets.sh: reproducible fixed-seed asset generation through
MODELBEAST (flux_local -> bg_remove_local -> hunyuan3d_mlx). All on-device.
Legibility work that mattered: butter rendered invisible at first. Fixes were
a gamma-style wetting darkening (a multiply *brightens* pale crumb), a normal
perturbation taken in UV units rather than raw texel deltas (~30x too small),
and pulling the lighting back under 1.0 so ACES stops desaturating yellow to
cream.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
