A Spore-inspired cell-stage evolution game written entirely in EigenScript — and the cells learn to play it themselves.
Eat, evolve, survive. Start as a single cell in the primordial pool. Consume food, avoid predators, spend DNA on mutations, call a mate, and choose when to evolve through five scale tiers.
Spore's creatures ran on scripted AI. In Tidepool the cells learn: a neural policy perceives the pool the way a player does — a 433-feature observation — and is trained with a from-scratch DQN to feed, survive, and time its evolutions. "Spore's cell stage where you watch the cells evolve their brains, written from scratch in a homemade language" is the pitch. There is no Python, JS, or native code here — the game, renderer, trainer, and editor are all EigenScript.
- Faithful cell stage — diet choice, eat food and meat, spend DNA on mutations and unlockable parts (mouths, fins, claws, poison/electric/jet), call a mate, evolve through five scale tiers, dodge giant "epic" cells.
- Evolution as a decision — eating enough food unlocks the next tier, but evolving is a deliberate action: each tier summons a harsher world (more, stronger predators), so when to evolve is a real risk/reward call.
- A learned brain — a 433-input MLP policy trained by an in-language DQN (replay buffer, target net, manual backprop). Press Tab in-game to let the policy pilot the cell.
- Bioluminescent renderer — translucent glowing cells, a depth-graded pool, caustics, and drifting plankton, drawn through a headless-inspectable draw-list layer so frames can be tested without a display.
EigenScript v0.19.0 or later — the neural policy stores its weights as flat
shaped buffers (#275, v0.19.0), and background music needs the audio_music_*
builtins (v0.18.0) plus libsdl2-mixer-2.0-0 at runtime. CI builds against the
pinned version in .devcontainer/Dockerfile's EIGS_REF. Check it out
alongside this repo and build it:
git clone https://github.com/InauguralSystems/EigenScript.git # ../EigenScript
cd EigenScript && make install-gfxThe Makefile builds the right binary from ../EigenScript (override with
EIGS_DIR=) and works from any directory:
make run # build the graphical binary and play
make test # build the headless binary and run the test suite
make lint # parse-check every source
make shot # regenerate the gameplay screenshot headless (needs xvfb)Or run directly: /path/to/EigenScript/src/eigenscript tidepool.eigs.
The repo ships a devcontainer (.devcontainer/) that builds a pinned EigenScript
from source, so it also opens in a GitHub Codespace
with the toolchain ready — run make test / make shot there without a local setup.
| Keys | Action |
|---|---|
| WASD / arrows | move |
| 1 / 2 / 3 | buy speed / sense / spikes (5 DNA each) |
| E | evolve (when eligible) |
| M | call a mate |
| Tab | watch mode — the AI plays |
| Esc | pause / back |
Training is headless (no display) and slow on an interpreter, so it runs off-line:
EIG=../EigenScript/src/eigenscript
$EIG train.eigs --episodes 800 --seed 42 # -> models/policy.txt (+ train_log.csv)
$EIG train.eigs --episodes 800 --resume # chain runs toward convergence
$EIG eval_policy.eigs --seeds 20 --ticks 1000 # trained vs autopilot vs random (mean ±95% CI)train_log.csv logs mean_tier per block — the signal for whether the policy
is learning to evolve. See CLAUDE.md for the full workflow.
New here? The good first issues are small, self-contained tasks with the exact file, the fix, and acceptance criteria spelled out — a good place to start.
Build and test with make test / make lint. The repo doubles as a real-world
EigenScript stress test — language gaps found during development are tracked in
GAPS.md and driven upstream. See CONTRIBUTING.md.
MIT — see LICENSE.