Browser Engine Lab
One production engine: train on massive local data, test search changes at fixed nodes, and keep heavy compute off the Pi.
Browser engine position
Analysis
Feature contributions
Weighted contributions before the score transform.
All features
| Feature | Value | Weight | Contribution |
|---|
Local training corpus
Disk-backed SQLite on your machine. Reuse the normalized desktop DB directly or build a new corpus from huge PGN files.
Connect the companion to inspect corpus coverage.
Import huge local data
Label corpus
Idle.
Active learning
Scan the training partition with the current candidate, find the largest model-vs-Stockfish disagreements, then relabel only those hard positions more deeply. Deep labels stay local in this browser and are weighted 3× during later training.
No active-learning overlay yet.
Stream training
The browser streams mini-batches from local SQLite. Game groups are deterministically assigned to train or validation. With a position cap, the trainer builds a phase-balanced, source-group-capped reservoir instead of taking the first rows.
Production Main uses 51 active evaluator features. The compatible model file still has 54 slots, but stmHasCheckSign, stmHasCaptureSign, and mateIn1Net stay zero and are not optimized.
Training quality
ReadyA lower evaluation error is not enough to publish. Use Evaluate for tactical/endgame suites, search-profile A/B tests, browser-engine head-to-head, and Stockfish Elo calibration.
Search performance benchmark
The frozen 2026-10-08 model (~1701 Elo estimate) plus Frozen Main · 2026-10-08 is the immutable pre-promotion baseline. Production Main now uses the fast evaluator with razoring disabled. Fixed-node and fixed-time passes alternate order to reduce JIT/order bias.
Per-position benchmark details
No benchmark run yet.
Main vs frozen baseline arena
Reproducible architecture comparison: frozen model on both sides, promoted Main vs Frozen Main · 2026-10-08, 20 distinct opening lines with colors swapped. Reports both completed-only scoring and the conservative score that treats every ply-limit game as a draw.
Promotion arena details
No promotion arena run yet.
Candidate vs baseline
The frozen Main baseline model is protected from deletion. Same-model comparisons are allowed when the profiles differ.
Profile A/B test
Same frozen model, same openings, same node budget. The default A/B compares promoted Main against the frozen pre-promotion Main.
Tactical / endgame strength gate
Quick regression gate for production Main. It runs the frozen ~1701 model through both Tactics and Endgames with promoted Main vs Frozen Main · 2026-10-08, alternating profile order by fixture.
Paired suite details
No paired suite run yet.
Manual single-profile suite runner:
Tactical pruning isolation
Diagnostic matrix for the fast evaluator. It reruns the two mate-in-2 failures plus five control puzzles while disabling one selective-search heuristic at a time. This is intended to identify the search heuristic hiding short forced lines before any production change is made.
Isolation details
No pruning-isolation run yet.
Lichess puzzle · Stockfish ranking
The fresh holdout seed 20261008 remains prefilled for reproducible post-promotion checks. Compare promoted Main against the frozen pre-promotion profile.
Puzzle ranking details
No puzzle benchmark run yet.
Stockfish Elo calibration
The candidate always plays through BrowserAlphaBetaEngine. Local Stockfish supplies only UCI-limited opponent moves.
Connect Stockfish to detect its supported Elo range.
Idle.
Local models
Keep candidates reproducible: train, run fixed suites and search/profile tests, estimate Elo, export, then submit.
| Model | Held-out MAE | Elo estimate | Baseline score | Tactics | Actions |
|---|
Owner review and publication
Loading access…