README.md

metadata

title: Eleusis
emoji: 🃏
colorFrom: green
colorTo: gray
sdk: static
app_file: index.html
pinned: false

Eleusis

Play Eleusis — the card game of scientific induction — against a hidden rule, and compare your game with a trained 4B model that played the exact same deal.

Two variants, switchable at the top of the page:

	🏆 Benchmark rules	🎓 Simple rules
Hidden rules	26 eval rules from the Eleusis benchmark	6 held-out simple rules × 4 deals
Turns	10	8
Redraws	played card replaced from the deck	none — your hand shrinks
Model replays	Qwen3.5-4B-eleusis-rl4-consistency-step30 (26×8 verification run)	Qwen3.5-4B LoRA, step 90 of this hosted RL run (24×8 eval)

How it works

1. New game draws a hidden rule from the selected variant's eval set.
2. Each turn you pick a card from your hand to test: the rule accepts it onto the mainline or rejects it.
3. At any turn you may also write a rule guess (a small Python predicate over card / mainline), or describe it in English and let ✨ Translate draft the Python. Guessing is free and judged by behavioral equivalence — the exact check used to evaluate the model.
4. Reveal the model's game whenever you like: real recorded games on the same rule and the same deal (starter card, hand, and draw order) — not live inference.

Fairness notes

• You and the model face the identical deal. In the benchmark variant, if you draw past the model's recorded sequence, remaining draws come from a per-rule deterministic shuffle.
• The model submitted a rule hypothesis every turn with no tools beyond play(rule, card) — the same interface you have here.

Built with the eleusis environment; the in-Space rule engine is extracted verbatim from it.