Evolution — ecosystem & evolution simulation

A deterministic Rust ALife engine + zero-build HTML/Canvas viewer, decoupled over WebSocket. Creatures carry neural-net brains whose weights live in their genome — behavior is evolved, not scripted. The ecosystem runs on a closed energy loop (sunlight in, metabolic heat out), so carrying capacity and boom-bust cycles are emergent. Populations isolated in different biomes face different selective pressure and diverge.

See plan.md for the full architecture and design decisions.

What's real

• Evolved brains — 7-ray retina + interoception → recurrent MLP (966 weights in the genome) → motor outputs (turn, throttle, eat, attack, mate). No hardcoded seek/flee.
• Diploid genetics — 13 body loci with dominance + recombination, haploid brain weights with segment crossover, evolvable per-site mutation rate (mutator locus), MHC-like immunity vector.
• Closed energy loop — soil → plant → herbivore → carnivore → carcass → soil. Sunlight is the only inflow, metabolic heat the only outflow. Total energy is conserved every tick (tested).
• Allometric metabolism — Kleiber mass^0.75 basal cost + locomotion + sensory + neural upkeep, modulated per biome (desert punishes sprinting, tundra costs homeostasis).
• Life history — birth → growth → maturity gate → sexual reproduction with egg provisioning; Gompertz senescence from an evolvable longevity locus (disposable-soma trade-off).
• Deterministic — one seeded RNG, fixed draw order, deferred births/deaths. Same seed → same run. Headless turbo mode fast-forwards; the WebSocket viewer is a read-only tap.
• Honest bootstrap — pure-random genomes, no safety net. When a population dies out the world auto-resets from a derived seed and increments a failed-attempt counter.

Biomes: Tundra · Desert · Rainforest · Volcanic (vertical bands).

Build & run

Requires a recent Rust toolchain.

Headless (evolve fast, print stats)

cd backend
cargo run --release -- --headless --seed 42 --ticks 200000 --every 20000

Prints pop / plants / carcasses / max-generation / resets | mean size per biome | total energy. Two runs with the same --seed produce identical results.

Live (watch it in the browser)

# terminal 1 — the engine + WebSocket server (ws://127.0.0.1:8787)
cd backend
cargo run --release -- --seed 42

# terminal 2 — serve the static frontend
cd frontend
python3 -m http.server 8000

Open http://localhost:8000. You'll see creatures move, eat, reproduce, and die; the sidebar shows live population/generation/energy and per-biome stats. Spawn Disaster kills a slider-set fraction of the population — watch the population dip and recover.

Tests

cd backend && cargo test

Covers metabolism, genetics (gamete/fertilize/mutation bounds), brain forward pass, senescence monotonicity, energy conservation each tick, determinism (same seed → same state), and extinction reset.

Tuning

Energy balance is the live knob — all rates live as named consts in backend/src/config.rs. The default regime shows large-amplitude grazer-plant boom-bust with occasional extinction+reset; that's honest emergent ALife, not a bug. Edit K_*, PLANT_GROWTH_R, reproduction, etc. for gentler or harsher dynamics.

Layout

backend/src/
  config.rs   consts (topology, metabolism K_*, energy rates)   rng.rs   seeded ChaCha8 + substreams
  vec.rs      Vec2                                                spatial.rs uniform grid + ray march
  genome.rs   diploid body + haploid brain, gamete/fertilize     brain.rs   MLP forward pass
  sensors.rs  retina + interoception                             biome.rs   biomes + soil substrate
  creature.rs metabolism, growth, senescence                     food.rs    plants + carcasses
  world.rs    deterministic tick + closed-loop ledger + reset    protocol.rs wire types
  net.rs      WebSocket viewer server                            main.rs    headless | live CLI
frontend/
  index.html  canvas + dashboard    main.js  WS client + render loop

Phase 2 (layered on the running core): seasons, phenotypic plasticity, sexual selection + mate choice, disease/parasites → Red Queen, click-inspect + lineage tree, SoA/rayon/binary transport.