OmniAstra

An experimental platform for testing whether one Transformer, trained across many structurally different card/strategy games, learns strategic representations that transfer to an unseen game.

• docs/DESIGN.md — architecture, decisions, limitations
• docs/EXPERIMENTS.md — transfer-experiment protocol and controls
• docs/RESULTS.md — what has actually been run, with numbers

Workspace (crates/): omnia-dsl (game IR), omnia-engine (deterministic simulator), omnia-gen (procedural game families + controls), omnia-observation (variable-length tokens), omnia-model (Burn Transformer), omnia-search (agents, ISMCTS), omnia-training, omnia-eval (matches, transfer runner, self-play), omnia-cli (binary omnia). Hand-written games live in games/ (including melee, a six-player free-for-all), experiment configs in experiments/.

cargo test --release --workspace
cargo run --release -p omnia-cli -- game generate --family tempo --count 3 --out /tmp/g
cargo run --release -p omnia-cli -- play games/skirmish.ron --p1 mcts:64 --p2 random --games 20
cargo run --release -p omnia-cli -- benchmark --model nano
cargo run --release -p omnia-cli -- transfer experiments/smoke.json