Test 01 · ran 2026-07-09

One prompt. Four models.
No winner.

Run all four, live Take the blind test

The verdict

For a task like this, the model doesn't matter.

This ran expecting a leaderboard — a best, a worst, at least one model that fumbled the math. That's not what happened. Four separate codebases independently implemented the same partial sum and converged on the same approximation error, within a rounding error of each other. The precise claim: What's left to optimize is cost, speed, and taste — and one of those diverged by roughly 200×.

The wall-clock times in the post feed the Speed & temperament measurement. The video plays on X — this site makes no external requests.

The prompt, verbatim (prompt.md) Raw results (JSON)

MSE = mean-squared error of the traced wave vs an ideal square wave, read off each app's own live readout. Four different authors, one rounding error apart.