A browser lab where tiny neural networks evolve real behavior — and nothing about how they behave is ever programmed.
Live: pouyanjay.github.io/darwinlab
Each fish carries its own 68-weight neural network. It is never told to flee the shark — most early fish don't, and get eaten. The ones that happen to swim a little less badly breed, mutate, and the next generation is born from them. Repeat a few dozen times and the tank is full of fish that dodge, wall-hug, and scatter — behavior that was selected for, not coded. The shark is fixed rules the whole time: it is a filter, not a teacher.
The five worlds on the bench are an ablation ladder: each gives evolution one more sense to work with, from nothing (Blind drift) to distance, direction, closing speed, and wall awareness. The result — measured headlessly over many seeded runs, not eyeballed — is that more senses is not more intelligence:
- Direction is the only sense that clearly pays, and only by a few points of survival.
- Distance alone barely beats sensing nothing at all.
- Closing speed and wall sense don't stack on top — this predator never creates dangers those inputs uniquely solve, so evolution never learns to use them. An input the world doesn't reward is just noise a brain has to overcome.
The lab refuses to fake a clean "more senses → smarter" ladder, because that isn't what happens. A nightly CI job re-measures the sweep and fails if the finding drifts.
The sense ladder shows inputs that don't earn their keep. The Shoal exhibit shows one that does — and the contrast is the whole point: a sense is worth exactly what the world makes it worth.
Switch the bench to The Shoal and two tanks run the same dangerous ocean and the same brain, differing only in whether the fish can sense each other. Give the predator a confusion effect — it can hold only one target, and loses its lock (then mills, distracted) when that fish is buried in a dense crowd — and schooling evolves on its own. Nobody writes the flocking; it appears because it works. Measured as the shoal sense's marginal effect (sense-on vs sense-off at the same ocean, over many seeded runs), fish that can feel their neighbours pack ~20px tighter and survive ~4s longer per generation. Toggle the shoal pill off and the school comes apart.
The honest caveats stay: cohesion (tight grouping) is the robust signal; alignment is weaker and lineage-variable, and the readout says so. And it only pays because the shark is fast enough that fleeing alone can't save you — slow the shark down and grouping stops mattering, exactly as the thesis predicts.
- Watch five worlds evolve side by side — live learning curves, real populations, every fish an individual genome.
- Ablate senses live: toggling a sense pill feeds 0 into that input slot of every brain in the tank — a true ablation on already-evolved minds, mid-swim.
- Open a mind: click any fish (★ Champion follows the best lineage) to see its actual evolved weights firing, what it senses right now, and its motor outputs — while it swims.
- Edit the world without wiping its learning: population, predators, mutation rate, arena.
- Train, then deploy: after the training horizon, evolution stops — no more respawns. The population the lab bred has to survive on its own, and you watch it decay (half-life included). Natural selection with the selection turned off.
- Switch exhibits: the sense ladder, or The Shoal — where an evolved swarm forms on its own and you can break it by switching off the sense that built it.
- Play the story: a narrated film of the whole argument, made of live simulations, not recordings.
Everything above is Studio — watching one world evolve. Flip the top bar to Research and the same engine becomes a console for running MANY simulations and reading a conclusion out of them: a left rail (the subject world, the instruments, the findings notebook), a workspace holding the active instrument, and a right sidebar of context. Batches run on a Web Worker pool, so thousands of bouts measure in the background while the lab stays responsive; nothing here edits the engine — Research only ever reads it.
The console is organised around the seven questions a rigorous study must answer, and each instrument is tagged with the ones it settles:
- The Sweep (Q2 · Q6) — pick some factors (each sense, predator speed, persistence…) and it runs the full factorial across your seeds, then reports each factor's effect on survival with a 95% interval. A bar that clears zero is a knob that matters; one that straddles zero is a knob that does nothing here — a kept negative, not a gap to hide. Intervals only, no significance badge.
- The Ledger (Q3) — state a claim ("direction pays more than distance"), and it designs the two arms, measures them, and returns a supported / refuted verdict from a single pre-registered contrast — the one thing that makes a verdict word honest.
- The Atlas (Q4) — choose two parameters and it paints a survival landscape you can pan and zoom: coral where fish die fast, teal where they last, with the cliff drawn down the column where survival actually falls off. Drill any point and Watch this world carries that config into Studio.
- The Trace (Q1 · Q5) — a second kind of study: it evolves ONE population keeping its genomes, reads the learning curve (did it climb and converge?), then traces that evolved school against a random-brain control on the same bout — the mechanism (accurate fleeing, distance kept) reads as the gap between the two, in bars and in the paths each school actually swam.
- The Report — the fifth instrument assembles the findings you keep into a seven-question brief: a headline, an answers-at-a-glance table, each question drawn from the evidence that settled it, and a method footer. Its honesty rail is the point — a question reads answered only when a real finding backs it; the rest stay honest "run the test" prompts. Export it as Markdown, print it to PDF, or Watch its subject back in Studio.
The two modes are one lab: Analyse on any Studio world hands it to Research as the subject every instrument then explores; Watch this world brings a finding or a map point back to the bench. The honest finding survives the harder look — Direction is still the only sense whose Sweep bar clears zero.
- Network: 8 inputs → 6 hidden (tanh) → 2 outputs (turn, thrust). The genome is the 68 weights; there is no other memory or logic.
- Evolution: fitness = seconds survived. Elitism, champion preservation, tournament selection, crossover, mutation. The mutation-rate slider genuinely drives drift.
- Predator: fixed rules only — cruise → aim → lunge, predictive interception, and persistence (it gets faster and wider-jawed the longer it goes without a kill, so no fish is ever permanently safe).
- Honesty gates: the engine is a bit-exact port of the original reference implementation (a seeded fidelity test drives both off the same random stream and asserts identical state), and the survival claims above are pinned by seeded tests plus a nightly re-measurement.
More detail in ARCHITECTURE.md.
npm install
npm run dev # the lab, at localhost:5173
npm run check # svelte-check
npm run lint # prettier + eslint
npm run test:unit # vitest — engine, stores, components, the fidelity gate
npm run test:e2e # playwright — builds and previews first
npm run bench:survival # headless science measurement (slow; re-run after ANY engine change)
npm run build # static build (BASE_PATH=/darwinlab for a GitHub Pages-shaped build)SvelteKit 2 · Svelte 5 runes · TypeScript · client-only static build (no backend). CI runs
check → lint → unit → build on every PR; on main the full e2e suite runs against the
Pages-shaped build before deploying it.
Generations are synchronized, fitness is one number, and a shark is three rules in a trench coat. The point is not to model the ocean — it is to make the mechanism of selection visible enough that you can poke it and predict what happens.