cursor-fable-5-thinking-high — Isla Genesis

Solution for life-sim: a browser-viewable 3D island ecology where three species compete, reproduce, mutate, and visibly evolve.

▶ Open it

Playable artifact: index.html — open it directly in any desktop browser (double-click works; Three.js is loaded from a CDN, so you need an internet connection but no build step, backend, or API keys).

If your browser blocks anything on file://, serve the folder instead:

cd life-sim/cursor-fable-5-thinking-high
python3 -m http.server 8000
# open http://localhost:8000/

A mirrored copy is published at galaxybrain.dev/artifacts/life-sim/cursor-fable-5-thinking-high.html.

Harness / model

• Harness: cursor
• Model: fable-5-thinking-high

What I built

Isla Genesis — a single-file Three.js artificial-life simulation:

• Seeded island generation: fractal-noise terrain with radial falloff, beaches, water, moisture-driven plant distribution, rocks and palms. The seed in the toolbar reproduces the entire run setup; 🎲 rolls a new island. Starting populations and per-island species trait baselines are also seed-jittered, so initial circumstances genuinely differ between runs.
• Three ecological roles with real mechanics:
  • Glimmers (grazers) eat plants that only regrow in daylight.
  • Stalkers (predators) ambush-hunt with a stamina-limited sprint, digest cooldowns, and pounce probabilities that depend on size, camouflage and whether the prey is asleep. Kill leftovers become carrion.
  • Skitters (amphibious scavengers) smell carrion at long range, eat tide wash-ups on the shoreline, can flee into shallow water where predators can't follow, and nibble plants inefficiently when starving.
• Heritable 7-gene genome (size, speed, sense, camo, nocturnal, fertility, hue) with per-gene crossover and gaussian mutation (8% chance of large jumps, flagged as ⚡ mutants in the event log). Genes are visible on the body: size scales the animal, speed elongates it, sense grows the head, camo/hue/nocturnal recolor it.
• Behavior: a utility state machine — flee > sleep > mate > eat > wander — driven by sensed neighbors via a spatial hash grid. The day/night cycle changes light, sky, stars and mood, gates plant photosynthesis, and scales each organism's activity and vision by how well its nocturnal gene matches the darkness (mismatched animals sleep and become easy prey).
• Instrumentation: HUD with clock, populations and sim health; click any organism to inspect genome, energy, age, generation and behavior state (with camera-follow); an Evolution tab with population history, per-species trait-drift charts, birth/death/generation counters and an event log (extinctions, mutants, generation milestones); a How it works tab explaining all rules in-app.
• Controls: pause/speed slider (0–16×, space to pause), seed input, restart-same-seed, and new-island buttons. Orbit/zoom/pan camera.

Typical 10-minute sessions (default 3× ≈ 22 island days, ~20+ generations) show grazer boom-bust waves against plant regrowth, predator-prey cycles, occasional extinctions or scavenger takeovers, and visible trait drift (e.g. grazer speed/camo creeping up under predation).

Implementation notes

• Everything is in one index.html (~1,700 lines). Dependencies: three 0.160 via CDN import map. No build, no backend.
• All organisms, plants, rocks and carrion are GPU-instanced; the terrain height field is cached into a grid for O(1) lookups; neighbor queries use a spatial hash; the sim runs on a fixed 1/30 s timestep with a per-frame step cap so high speeds shed load instead of freezing the tab.
• Population hard caps: 340 grazers / 240 scavengers / 48 predators.
• window.sim is exposed in the console for the curious (e.g. sim.history, sim.causes for death-cause tallies).

Evaluation pointers

The best file for evaluation is index.html. Suggested pass:

1. Open it, watch the HUD populations move; drag to orbit, scroll to zoom.
2. Click an animal → genome/energy/state panel; tick "follow with camera".
3. Open Evolution → population waves, trait drift, mutant/extinction log.
4. Push the speed slider right (8–16×) and watch a few generations turn over; compare trait averages after ~5 minutes.
5. Hit 🎲 New island — different terrain, plant layout, founder traits; the seed field reproduces any run.

Notes / limitations

• Needs internet once to fetch Three.js from jsDelivr (the only external dependency).
• Ecology is tuned for drama rather than equilibrium: extinctions are real and permanent within a run (a low-density budding fallback exists but a starving last survivor can still die). Different seeds produce meaningfully different outcomes — that is intended.
• Determinism is per-step, not per-wallclock: the same seed replays the same island and founders; long runs diverge slightly with frame timing because step counts per frame vary.