VEKELETH
The state of holding more finished futures than can ever be lived.
Everything below this line is running on your machine, right now. The search brain is written in freestanding C — integer arithmetic only, no standard library, no floating point, no allocation, and nothing learned — compiled to WebAssembly and handed to your browser as 23,992 bytes. Each frame it imagines a hundred and seventy complete futures, draws them, scores them, keeps one, and throws the rest away. The faint lines are the discarded ones. You are watching vorka happen ten thousand times a second.
BLUE = the search brain, 0 rollouts × 0 ticks of lookahead. RED = the greedy brain, 12 rollouts × 0 ticks. Same world, same seed, same compiled function, same physics, same scoring. The only difference between them is how far ahead they are allowed to look. Faint red lines are futures that hit a wall. Faint blue lines are futures that survived to the horizon. The gold line is the one it kept — for six ticks, until it is dropped too.
What it actually proves
A demo that only ever shows you the good run is an advertisement. So before this page existed, the same WebAssembly module was run headless over four hundred consecutive seeds, both brains, to termination or 1,400 frames. Here is the whole distribution, unedited.
Look closely at the eighteen seeds where the greedy brain won. On seventeen of them it won by between two and nine units — both brains died at the same wall, and the greedy one happened to be a nose in front when it hit. There is exactly one seed in four hundred — seed 138 — where the greedy brain genuinely outperformed: 4,909 against 1,695. That seed is in the rotation above. If you sit here long enough you will watch the search brain lose, and nothing on this page will stop you seeing it.
And the search brain dies too. It dies on 399 of 400 seeds. Looking further ahead does not make you immortal; the corridor keeps narrowing and the gates keep closing. It makes you last, on average, ten times longer. That is the entire claim.
What this is not
The 88 published speedruns drive real NES and Atari emulators. This is a purpose-built world written specifically so the mechanism is visible: a cave, a ship, gravity, and walls with one hole in them. It is the same method, not the same target. If you want the method against real hardware, that is on the replays.
There is no network, no weights, no training, no gradient, no dataset. The brain has never seen this cave before the moment it enters it. It knows only the rules, and it is willing to imagine dying eight thousand times per frame rather than do it once.
The value of an imagined future is x·9 + gold·5200 + survived·30 (+ room/3, −420000 if dead). I chose those numbers by hand and tuned them until the search brain stopped being a coward. That is a human prior, and pretending otherwise would be a lie.
The greedy brain calls the identical search() function on the identical physics with the identical scoring. Two integers differ: depth 96 vs 6, rollouts 170 vs 12. Nothing else. You can widen its horizon and it becomes the other brain.
No WebAssembly toolchain existed on this machine. There is no clang, no emcc, no wasm-ld. So a compiler was fetched, the forward model was written from nothing as integer fixed point in Q8, the search was written on top of it, the whole thing was compiled to a freestanding wasm32 target with no runtime and no imports at all — it asks the browser for nothing — and then it was benchmarked over 400 seeds before a single pixel of this page was drawn. The first version was wrong: without gates the corridor was survivable by pure greed, and the search brain proved nothing. The gates exist so that the experiment can actually fail.