Changelog

What shipped, when

2026-08 — Corrected: the site no longer promises to stay free

“Free” removed from the description of the project. The engine is open source and that has not changed, but the site had also promised the service itself was free, which is not something we can commit to: hosting, bandwidth and compute cost money, and a small fee may eventually be needed to cover them. Terms of use now says plainly what it costs today — nothing — and that a change would be announced here first.

Corrected claim: the site said “the engine that plays you here is the one you can download” and that make play plays you against Raccoon. The trained weights are not in the engine repository, and make play without a checkpoint quietly starts an untrained network. How it works now says so, and points at the ONNX file this site actually runs, which is published in the website repository. Wiring the Python engine up to load it is part of milestone 5.

Checkers no longer overlap. A point is now exactly five checker diameters long, the way it is on a real board, so a stack of five sits edge to edge instead of each checker hiding a slice of the one below it. Checkers gained the turned rim a real one has, and the board is drawn taller and wider to pay for the room.
The dice landed on top of the checkers. They now sit in the clear band between the point tips, on the roller’s side of the bar. In the opening position they had been covering the fifth checker of the six point.
Point numbers moved onto the frame, where nothing competes with them, and are large enough to read.
A sixth checker on the bar is visible again. The bar was capped at five with no count, so anything beyond that was silently not drawn; it now carries the same numeral the points do.
Raccoon’s turn can be followed. Its dice land, and its move appears a beat later, instead of the whole turn arriving in one frame. A line above the board alternates between the two sides — “Raccoon rolled 54: 13/8 24/20”, then “You rolled 51: 13/8 24/23”.
The ranked-move table moved under the board and shows the network’s full six-outcome split — W, Wg, Wbg, Lg, Lbg — in the cumulative convention XG and GNU Backgammon use. “Loss” is now “Diff”, and the heading over the table says “Raccoon ranking of moves” rather than asking the reader to guess.
The tab icon is a raccoon rather than a placeholder triangle.

The site was restructured around six pages so that a first-time visitor can see what Raccoon is, why it exists and how strong it is without reading the research notes. All strength numbers now live on one Strength page, and the development plan is written down on Roadmap.
Corrected number: the head-to-head result against GNU Backgammon is +0.047 ± 0.003 points per game over 18,000 games — the pooled figure across both seed families. The site previously quoted +0.046 ± 0.005 over 6,000 games, which was one arm of that measurement.
The “compute per decision” column was dropped from the strength tables. The seconds quoted in it were never measured; the tables now say what search each engine does, which is the comparison that was meant.

The exp018-ep22 network runs client-side via ONNX Runtime Web: play a full cubeless money game, or analyse any position, with no server and no account.
Analysis panel showing ranked candidate moves with cubeless equity and the win/gammon/backgammon split from the six-outcome value head.
XGID import and export, so positions move between Raccoon and existing backgammon tools.
The exported network was re-scored on the BGSage benchmark to confirm the browser engine matches the measured one (PR 0.9500 vs 0.9498, n = 14,693).

On a double, the engine was choosing its four half-moves two at a time rather than optimising the turn as a whole. It cost 0.062 points on the one doubles turn in six where the difference mattered. Fixing the search — with no change to the network — took the margin over GNU Backgammon at 0-ply from +0.013 to +0.046 points per game. How it was found. The browser engine plays the fixed way.

Variance-reduced rollouts put exp018-ep22 ahead of GNU Backgammon at 0-ply over 6,000 games — the first time a Raccoon network has been shown to beat the teacher it was distilled from in actual play. The plain average over the same games settles nothing at ±0.035; the variance reduction is worth about 55× as many games. Method and result.

Scaling 2-ply distillation from 8M to 40M positions cut move-selection error from PR 2.16 to 0.93 and brought value accuracy to teacher level, with no trade-off across game plans. Write-up.

Every experiment behind these releases is published in full, including the ones that failed, at research.raccoonbg.com.