The concept is the contract

The platform's concept-image service was out of credits this cycle (credit_balance_exhausted, logged in concept/gen.log), so the three concept sheets were painted in-lab instead — which turned out to be the stronger contract. Every colour on them is a named hex and every proportion is a stated measurement, so the Blender build and the shipped file are checked against the sheet numerically rather than by eye. The gorge sheet below fixes the three parallax bands, the mist fall-line and the rule that the sky ramp IS the clock.

Concept sheet 2/3 — the gorge, its depth bands and the dusk ramp

The same gorge, live in Phaser 4 at 42% dusk

Gray boxes before art — and the level placed on measured trajectories

The whole gorge ran, and was crossed by an automated player, as gray boxes before any art existed. The feel layer was tuned there: the rope constraint, the pump, the reel, the two gravities, the speed-coupled camera. Then a reachability probe drove the shipped step() directly — thousands of pumped-swing-and-release trials with no rendering and no wall clock — and reported that a pumped swinger rises 202–558 px above its own ring and carries up to 2353 px forward. Every nest and every spur ledge in the level sits on a point one of those trials actually reached; closest approach is 0–10 px for all six. Nests are not placed by eye in this piece.

Blockout: the complete feel layer, no art

Phone landscape — hold the right half, drag the left

Code-modeled in Blender, animated from real motion capture

The gatherer is a code-built model — no sculpting by hand, no asset store. The swing cycle is motion capture, but not from one clip. The studio's curated library had no hanging motion at all, so the full CMU archive's own description index was read rather than guessing subject numbers — which found subject 43, "playground: grip bar, swing body". Retargeted and then measured frame by frame (wrist height against head height, torso verticality), that clip holds a usable hang for eight frames out of 181, and eight adjacent frames at 120 fps are the same picture: a cycle sampled from them fails an adjacent-frame difference gate, correctly.

So the cycle is composed. The arm chain is held at an overhead grip and legs and spine are driven by a locomotion clip (CMU 02_01), with the arm F-curves stripped out of the bake — because a keyed bone is re-evaluated from its action on every frame change and silently overwrites a pose set by hand. Arms locked on the line, legs scissoring fore-aft, and an authored hip sway and bob on top: which is what pumping a swing looks like. The shipped sheet passes the adjacent-frame gate at 6.17–9.70 against a threshold of 4.0 — it failed at 3.81 first, and the fix was to add the missing motion, never to lower the gate.

The face is not baked into the body sheet. It renders as its own five-state sheet — neutral, blink, alert, win, hurt — and the pipeline exports the head pivot in frame pixels for every body frame, so the engine composites the face onto any pose. That is what lets the character blink every few seconds and change expression mid-swing without multiplying the body sheet by the number of expressions.

The gatherer — studio side render, round 4 of 4

Concept sheet 1/3 — the contract the render is sampled against

The shipped swing sheet — 8 frames at 112 px, mocap-driven

The five face states, each cropped to its own ink and enlarged 11× — at shipping scale they are a few pixels each

The prop kit as shipped — every named thing on the concept sheet is a real file

…and the sheet those props were authored against

The feel research

This lab's output is not only the game — it is the numbers and the reasons. Every constant was A/B-swept against the shipped simulation, and the values that were thrown away were measured, not asserted. Two of the sweeps:

The reversal count is the readability metric: a pendulum reverses, a motor does not. The second sweep is the one that surprised us — reeling the line in conserves angular momentum, so pulling yourself up the rope at the bottom of the arc genuinely makes you faster, but only up to a point.

Measured

Frame costs are from the audit box: a shared, GPU-less host running Chromium's software canvas renderer. That is the worst case this piece will ever see and the numbers are reported as measured, including the ones we would rather not print.

The unflattering row. 60 fps is not demonstrable on this audit box: it has no GPU, so Chromium falls back to a software canvas rasteriser and the frame cost we measure is dominated by rasterisation, not by the game. On real hardware the same build runs the WebGL renderer. We report what we measured rather than what we expect.

What the piece earned

From concept to rooftop

The piece was concepted first: a key-art sheet, a character turnaround and an environment callout board set the palette — terracotta caps, cream plaster, white linen, a slate-blue storm. Every named element in the sheets was then authored as a real asset.

Concept key art — the contract the piece is audited against

The same morning, live in the engine

Gray boxes before art

The whole game ran and was beaten as gray boxes before any art existed — feel (coyote time, jump buffering, asymmetric gravity, camera lookahead) and the level's fall-recovery routes were tuned where changes cost minutes. An automated player drives every build through all eight lines to a verified win on each audit round.

Blockout: the full level playing as gray boxes

The storm at 90% — rain, tinted world, meter nearly out

A Blender pipeline, rendered to sprites

The heroine is a rigged Blender character whose run and jump come from retargeted CMU motion capture, rendered to sprite rows. The washing lines are real cloth simulations — three wind phases baked into a flutter flipbook. Tiles, props, skyline ranks and the storm bank are all Blender renders packed into Defold atlases.

The instant-load signature

This lab exists to prove a real-engine game can be playing before a phone user loses interest. Every payload ships pre-compressed and is inflated in the browser, so any static host serves it optimally. Measured first-playable times, Playwright CDP network throttling:

Total transfer 1.82 MB (engine 976 KB + game data ~0.8 MB, gzip + RDO-basis textures). Measured headless on software rendering — real devices boot faster.

Real engine features, earned