Best ladder Elo
-
last published run

The scorebook

The engine trains nightly by gated self-play with human-game replay, and Stockfish signs the rating. Each position is handed to the agent that owns it:

MDP · exact endgames PPO · on-policy RL Reward · DQN, shaped Neural · self-play net
-
Peak Elo
-
Generations
-
Centipawn loss
-
Top feature MI (bits)

Training pipeline

01Pre-train

Behavioural cloning on real Lichess games.

pretrain/model.pt
02Post-train nightly

Gated self-play with human-game replay. Promoted only on a win.

posttrain/gen-*.pt
03Serve and grade

The champion, scaled to an honest rating; Stockfish grades it.

/move · /calibrate

Strength over training

Each point is a checkpoint, rated by real games (brass = the per-checkpoint measurement, small-sample noise; claret = best reached). The curve is anchored so the endpoint equals the live Stockfish-calibrated rating, and nightly training appends new entries.

How the Elo is earned

Stockfish is throttled to known Elo bands and the agent plays gauntlets against each; a Bradley-Terry fit turns the scores into one calibrated rating. The agents never use Stockfish to choose a move - it only grades them.

Information theory

Which board features carry the most information about who wins - published with the analysis run.