Action-space decomposition
The flagship result, made interactive: a flat primitive-move search explodes super-exponentially on a rule-assembly puzzle, while an LM that proposes macro-actions collapses it to a single simulation.
On a BFS-certified ladder — push the WIN word to form FLAG IS WIN, then reach the flag — the FLAT searcher's median sims-to-solve grows 7 → 23 → 469 → 1372 and begins to censor at distance 6, while the MACRO arm (each macro grounded by a bounded local BFS) holds 1 sim and 100% solve at every distance. The reward is a gradient-free plateau ending in a one-cell needle; the macro injects the structure that removes it (whitepaper §decomposition).
Play it
A compact build of the same engine, live in your browser (Rust→WASM) — press Solve · FLAT vs Solve · MACRO on a hard rung and watch the sims diverge:
A trimmed embed of the explorer (canvas · decomposition · controls). Open the full app → The WASM bundle is staged by scripts/bundle_demos.sh before deploy.