Roadmap
Five milestones, each with the opponent it has to beat
Nothing here is called done until it has beaten a named opponent on a metric fixed before the run, at a sample size that makes the result mean something. One milestone has cleared that bar. The other four have not.
1. A network stronger than GNU Backgammon at 0-ply — Done ✅
A residual network, trained by distilling 40M positions labelled with GNU Backgammon’s 2-ply search, that beats GNU Backgammon’s own network at equal search depth on cubeless evaluations.
Measured two independent ways: PR 0.95 against 2.14 on the BGSage money benchmark, and +0.047 ± 0.003 points per game across 18,000 head-to-head games. See the evidence.
2. Search on top of the network, past GNU Backgammon at 2-ply — Next
The current engine does no search at all. Adding one — the hard part in backgammon being the chance nodes between every pair of decisions — should be worth more than any further network training.
The bar: BGSage benchmark PR below 0.56, GNU Backgammon’s 2-ply score, on the same 14,693 decisions.
3. A cube model better than the Janowski formula — Planned
The doubling cube is half of competitive backgammon and the engine does not handle it at all today. Cubeful equity is currently approximated everywhere by the Janowski formula with a fitted cube-life index; a model trained to predict it directly should do better.
The bar: lower cube error than the Janowski implementations in current engines, on money-game double/take/pass decisions with rollout references.
4. Match play — Planned
Everything above is money game. Match play changes the objective: equity is measured in match-winning chances, which makes both checker play and cube decisions score-dependent.
The bar: match-equity-aware evaluation and cube handling, benchmarked on recorded match decisions rather than money-game ones.
5. The engine behind a REST API — Planned
The board on this site is deliberately basic, and a polished interface is not something we plan to build. The better answer is to expose the engine cleanly and let people who build interfaces build them: positions in, ranked plays and equities out, over HTTP, with no Python to install.
The bar: a documented, stable, public endpoint that a front-end developer can build a real product against. If that is you, tell us what you need from it — this milestone is much easier to design with a first consumer in the room.
For what has already shipped and when, see the changelog. For the experiments behind each step, including the ones that did not work, see research.raccoonbg.com.