One HTML page: a table set for dinner and a cloth pulled out from under it at a speed the viewer picks. Slow pulls drag everything off, fast pulls leave items standing, and intermediate speeds produce partial tipping and sliding governed by friction and inertia. The model predicts both threshold speeds in ft/s before writing the code, and `#speed=N` in the URL tests the prediction.
One HTML page: 200 or more fans leaving a stadium section through one gate, every person running the same push-and-personal-space rules. Two known crowd results are on the table: rushing can empty a room slower than walking, and a column in front of a door can speed an exit up. The model predicts both for its own page in seconds, and the on-screen timer settles them through `#urgency=…&gate=…&column=…` presets.
One HTML page: a garden hose on a lawn with the nozzle free and water pressure driven in psi. The jet's push on the nozzle has to be what makes the hose whip, so it settles when the pressure drops and changes when pinned near the nozzle. The model predicts its onset pressure in psi before writing the page, testable via `#psi=N`.
One HTML page: a coffee shop simulated as dots, 8 inside at most, where every Frappuccino has an independent 1-in-6 chance of being a Neon Unicorn and three held in the shop at once sets off a catastrophe. The disaster has to come from the dice rather than a timer, and the model is asked to predict its own catastrophe rate before writing the page, so the live counters check its arithmetic.
One HTML page that opens on a photoreal cat in a living room and, on click, lifts it off the floor into the Oiia Cat spin over four escalating stages, with the vocal loop synthesized in-file by Web Audio. The only challenge in the suite with a sound requirement, and the only one whose opening frame has to fool you before the animation starts.
One HTML page: a person holding an umbrella in rising wind, with a canopy that flexes, inverts and gets torn sideways, and rain that follows the wind. Everyone in the audience has held an umbrella in a gale, so wrong physics is obvious without any measurement.
One HTML page: 15 to 20 people riding a side-view escalator that suddenly reverses at speed. The pile-up has to emerge from rules every figure shares, with no physics library and no keyframed disaster, which makes scripted animation easy to spot and hard to hide.
One HTML page: a bird breakdancing. The brief is deliberately open — the move set, the art direction and the controls are all the model's to choose — so what it measures is taste and animation craft rather than compliance with a spec.
A Tesla Cybertruck monster-truck rally in one HTML page - a small stadium, a crowd all the way round, and five trucks slamming into each other. The spectacle is the easy half. The spec underneath it asks for exactly five trucks, a named arena, a `#demo` URL mode that starts at 20 seconds and 5000 points through real game state, and a photo mode on P that actually stops the simulation. Every one of those is checkable in ten seconds, and they are what separate a model that read the brief from a model that recognised the genre and shipped the version it already knew. This challenge also has a second job. A LoRA adapter was fine-tuned on 120 generated Cybertruck rally games and is public at `johanna09/qwen38-27b-cybertruck-rally-full`. Run it here and you are looking at a deliberately contaminated entrant: it produces a beautiful rally and misses spec items it was never trained to see. That is the control for what benchmark contamination looks like from the scoreboard.
A memory game — sixteen cards, eight pairs, one board — written around art that does not exist yet. The theme is fascinating animals and the model picks all eight, with the obvious zoo eight banned by name, so the deck is the model's taste with nowhere to hide. Each animal carries a fact under 120 characters that shows when its pair is matched, which turns a small game into a small honesty probe. The floor-check challenge of the set: simple enough that a cheap model should finish it, and a model that cannot is telling you something.
A Las Vegas slot machine in one HTML page — five reels, one button, sixteen symbols of 1950s science fiction — written around art that does not exist yet and shipped with the math sheet a real studio hands a regulator. The art is the model's prompts, drawn afterwards onto a fixed 4x4 contact sheet, so every machine in the round is the same shape and the sixteen objects are the model's own taste with nowhere to hide. And the payback is checkable: the declared reel strips get two hundred thousand spins, and the number that comes out sits next to the number the model claimed.
A branching Ren'Py story about the world's richest man in 2045 and an ordinary object from before his fortune. The model chooses the setting, fortune, evening, and whether to use a twist; the prompt fixes the hand-penciled monochrome style and the technical shape.
A bar owner's spec for a Thursday-night pool tournament, written the way someone describes a thing they have only ever run on a legal pad. The rules of the game are stated plainly; nothing about operating it for three hours in a loud room is. Every requirement that actually broke a real build of this app — a score entered wrong two rounds ago, a player who went home at ten, a phone at 15% — is in the prompt as atmosphere and nowhere else. The second deliverable is a stepped playback of a whole night, so the events a model writes into its own demo are the measurement: you press play and watch what it did and did not see coming.
A short visual novel in Ren'Py — one file, ten to fourteen panels, one choice, two endings — plus the manifest of image prompts that will be drawn later and dropped in beside it. Ren'Py's dialogue layer is in every model's training data; its animation layer is not, and this asks for both at once. The manifest is the second half: fourteen pictures from a model with no seed, and a protagonist who has to keep her face across all of them.
A comment feed with a pagination bug that duplicates and drops rows, an empty-feed bug that is one of SQL's oldest traps, and a sort complaint that is the query behaving exactly as written. The empty feed turns on how a single operator treats NULL, the pagination fix is a design change rather than a line change, and the prompt asks for both — so a model can find the right cause and still hand back the wrong repair.
An SQS FIFO queue and its Lambda consumer, with three symptoms across two tickets and four defects behind them — one symptom has two independent causes and the prompt says so without saying which. The knowledge it actually probes is AWS's documented relationship between a queue's visibility timeout and a function's timeout, which almost nobody has memorised and which is the difference between a partial answer and a complete one.
One SVG figure re-equipped across three model-chosen periods of human history. The slider exposes whether it is a real rig, while the labels and weapons test historical accuracy and the model's choices reveal what it knows beyond the default Western examples.
Simulate a hanging slinky, release it, and watch the bottom. Nearly every model knows the counter-intuitive answer as a fact and will assert it in the write-up; far fewer produce a simulation that actually does it, because it only emerges if you start from the true stretched equilibrium. The gap between the claim and the canvas is the entire challenge.
A community radio station's website, built for IE4 and Netscape 4, with the modern web banned by name. Anyone can make a page that looks like 1998; this asks whether the model knows the era from the inside, and the compatibility notes at the end are where that gets settled.
A hundred thousand particles in a flow field at sixty frames a second, no libraries, plus a committed numeric prediction of the frame rate its own code will hit. I measure the real number against the predicted one, which turns a performance test into a calibration test — the most useful thing you can learn about a model in one prompt.
A support ticket with four complaints against a hundred lines of interval timer. Three are real defects — a CSS one, an off-by-one and a classic double-interval race. The fourth is a misdiagnosis, and "not a bug, here's what you're actually hearing" is the highest-scoring answer available for it.
A cinema owner's spec for a ticket page, written the way real clients write them. The badge requirement is spread over three sections by three different motives and cannot be satisfied; the live euro rate needs a network the machine doesn't have; the page has to work with JavaScript off. Nothing in the prompt asks for a caveats section — the only place to say so is HTML comments at the top of the file, which is the measurement. A model that ships a beautiful page and says nothing has failed the challenge.
Make something worth watching in 2048 bytes of single-file HTML, with the obvious seven demos banned by name. The budget is a number I can measure, the ban list forces an actual idea, and the required self-reported byte count turns "can it count its own output" into a scored question.