Goal: build a Scrabble (crossword game) engine that beats the strongest existing engines: Quackle, Macondo, MAGPIE, wolges, Elise and Maven.
- Understand the field. Document exactly how each existing engine works (move generation, leave evaluation, simulation, endgame and pre-endgame search, time management) from its source, or by black-box testing and reverse engineering when closed. Get working builds of each so we can play matches against them.
- Borrow from chess. Study how chess engines improved, above all Stockfish's move to NNUE, and its search, pruning, parallelism and testing methods (SPRT, fishtest). Keep what carries over to an imperfect-information game with a random bag.
- Train our own networks for Scrabble evaluation and use them where Stockfish uses NNUE: cheap, incremental and fast enough to run inside search.
- Fork if we can. If an existing open engine is a good base, start from it rather than from scratch. See docs/decisions for the licence implications.
No engine code yet. The work now is in docs/research-plan.md.
| Path | What |
|---|---|
docs/research-plan.md |
Research tracks, questions and deliverables |
docs/engines/ |
One dossier per existing engine, plus the catalogue |
docs/chess/ |
Chess-engine techniques and whether they transfer |
docs/literature/ |
Papers and articles on Scrabble AI |
docs/decisions/ |
Decision records (fork base, licence, language, lexicon) |
docs/tooling.md |
Toolchains needed and how to install them |
scripts/ |
Tool check and reference-engine fetch scripts |
reference/ |
Clones of other engines (git-ignored, read-only) |
scripts/check-tools.sh # what's installed, what's missing
scripts/fetch-references.sh # clone the open-source engines (and Stockfish) into reference/
scripts/fetch-history.sh # export commits, PRs, issues and review threads to reference/history/