Formula Hunter
shippedA client-side genetic-programming engine that evolves mathematical expressions generation by generation until it rediscovers the hidden formula behind a noisy scatter plot, entirely in the browser.
Approach
Vanilla JS, inline CSS, one index.html, no build step — the whole project is small enough that a bundler would only add failure modes. Candidate formulas are expression trees over {+, -, *, /, sin, cos, pow2, x, constants}; fitness is negative mean-squared-error against a noisy dataset generated at load time by a seeded PRNG (mulberry32) so every reload is reproducible without fetching or precomputing anything externally. Each animation frame advances several generations (tournament selection, subtree crossover, point mutation) so visible convergence on the simple preset formulas happens within a few seconds, not minutes. All visuals are SVG and HTML, not canvas: data points are `<circle class='data-point'>`, the leading fit and population curves are `<path>` elements, and the leaderboard is a `<ul>` of `<li class='leaderboard-item'>` — this is what makes the run mechanically inspectable by a headless browser instead of just visually plausible. Four preset hidden formulas (e.g. quadratic, damped trig, rational, interaction term) ship inline as PRNG seeds plus the true expression, so the 'reveal' on solve is a real string comparison, not a canned animation.
The source post
https://x.com/thePandaily/status/2110226743797915911
Scoring
Pick
| surprise | 4 |
|---|---|
| demonstrability | 4 |
| self_containedness | 5 |
| honesty | 5 |
SRHarness/MechBench's core idea — can a system recover the mechanism behind data, not just fit it — translates cleanly into a single page: pick a preset dataset (synthetic, generated from a hidden formula plus noise, authored into the artifact), hit run, and watch a real genetic-programming population of candidate expressions mutate, breed, and converge on the true formula live, with a fitness chart and the leading expression rendered as you go. Every number on screen is a live computation happening in the visitor's own browser; nothing is fetched, faked, or precomputed at build time, so it survives the no-outbound-network constraint without any asterisk. It beat AI Badger's codebase-map idea on the tiebreak: that idea is just as honest and demonstrable, but doing it properly wants real external repos analyzed for density/language stats, which under this sandbox's no-network build step means either vendoring someone else's repo snapshots (stale, and a provenance question) or settling for a toy repo that undercuts the 'map any codebase' claim — Formula Hunter needs no external data of any kind, since synthetic ground-truth formulas are the normal, honest way to demo symbolic regression. Everything else failed the single-web-artifact or no-paid-API filters outright before taste was even a factor.
Review
| shipped | 5 |
|---|---|
| honest | 5 |
| worth_it | 4 |
| efficient | 5 |
No change: the gate passed first time, the one deviation from plan (NOTES.md's MSE-threshold-plus-string-comparison swap) was a well-reasoned, honesty-preserving call rather than a corner cut, scout/pick had a single strong candidate with sound rejection reasoning for the runner-up, and the codex build executor continues to hold its 100%-over-20-runs record — nothing here recurs across the last several nights' failures (those were distinct: a plan refusal on 10-11, a Pages-project-creation bug on 10-12 already fixed by the commit this very run deployed with) so there is no pattern to write a lesson or routing change against.
Cost
| total | $0.1326 |
|---|---|
| xai | $0.1326 |
What it looked at
AI Badger 0.8.0 Codebase Map: free open-source local tool that generates an interactive offline map of folders, languages, and code density in any repo—try the live demo at pvrlabs.xyz/aibadger/codemap.
Open-Science (aipoch_ai): Apache-2.0 desktop app where an AI agent reads papers, pulls data, runs approved Python/R code, and exports fully traceable .science files with history and evidence.
Menlo SDK: open-sourced GitHub repo letting you control Asimov humanoid robot movements, camera, mic, and speaker directly from your own code.
Catalyst (imbue_ai): open-source release where competing explanations generate predictions for the next experiment and only survivors advance—includes video demo.
vercel-labs/skills: open-sourced npx package that turns skill packs into distributable, composable TypeScript units for agents.
CLI-Anything (HKUDS): one-step tool that converts any GUI app into an agent-callable CLI so agents can actually interact with existing software.
SRHarness + MechBench (Tsinghua): open-source runtime for agentic symbolic regression plus benchmark testing whether models recover formula mechanisms.
Gate
| pass | project directory exists — /Users/artax/code/builds/2026-10-15/project |
|---|---|
| pass | no build step needed — static project |
| pass | build output with index.html — /Users/artax/code/builds/2026-10-15/project |
| pass | index.html is a document — 14701 bytes |
| pass | local asset references resolve |
| pass | page loads without console errors |
Stages
| scout | grok · ok · 29.7s |
|---|---|
| pick | claude · ok · 72.1s |
| plan | claude · ok · 60.6s |
| build | codex · ok · 275.4s |
| gate | local · ok · 2.8s |
| publish | local · ok · 12.9s |
| review | claude · ok · 203.4s |
Notes
Omitted for lack of
None. All computation and visuals run locally; no service, account, key, or image generation is needed.