ENGINE RESEARCH · OCTOBER 2026
Pieces, learning, and the pawn oracle.
The pawn-only engine remains intact. Positions with knights, bishops, or rooks now use a tuned expert evaluator plus a small learned correction. Everything used during play runs on your device.
A small correction with a strong starting point
The new network has 29 inputs, 16 ReLU hidden units, and one output: 497 neural parameters. It reads pawn structure, material, activity, threats, promotion control, piece totals, and a projected pawn score. It adjusts a tuned 25-coefficient expert by at most ±2 evaluation points. Rules, legal moves, promotion replies, and searched wins remain authoritative.
This was selected after testing raw-board networks from 3,921 to 62,721 parameters. The largest model is 488 → 128 → 1, since 488×128 + 128 + 128 + 1 = 62,721. It still lost 36 games to the optimized expert’s 6 in its final native screening test, with 5 draws and 1 unresolved game. Extra capacity did not compensate for limited training data.
Can the pawn engine be an oracle?
It can supply a useful candidate feature: remove the extra pieces, query the existing frozen pawn network, and cache the answer by pawn structure, mover, and en-passant availability. The answer is clipped and lightly weighted. It is never an exact verdict about the real board.
Pieces change both tactics and tempo. A knight can make waiting moves that do not alter the pawns, so a pawn-only zugzwang may disappear. For example, White Pa7 against Black Ra1 and Ph2 loses after a8=Q Rxa8, even though the pawn-only projection wins for White.
The isolated oracle-on/off test scored 21 wins, 21 draws, 18 losses, and 4 unresolved games. That does not demonstrate a reliable independent gain. Searching the projected pawn game more deeply was too expensive and performed worse; an oracle-based move-ordering hint also showed no clear gain. The reliable handoff remains the actual transition to a pawn-only position, where the original engine and exact tables take over.
What the training budget bought
1,200 self-play games, about 1.07 billion search nodes, and 27,394 distinct positions. Generation consumed about 2.05 summed worker-hours on the available CPU environment. The small residual fit took 5.8 seconds. Producing the targets was the expensive part; no paid GPU service was used.
Whole games stay within one training or validation split, and symmetry-aware deduplication removes repeated positions. Of the games, 210 were unfinished at the experimental cap. Their positions receive lower-weight searched targets, never invented draw labels. Played outcomes and search targets are fallible estimates.
Fresh playing tests
Matched starting positions, colors swapped, equal thinking time, and the actual JavaScript search. The first three rows use 120 ms per move and randomized positions. The last two use the normal 800 ms budget and starting presets.
| Comparison | Wins / draws / losses / unresolved | Score on resolved games |
|---|---|---|
| Hybrid vs optimized expert | 46 / 28 / 29 / 17 | 58.3% |
| Hybrid vs previous mixed-piece evaluator | 74 / 16 / 18 / 12 | 75.9% |
| Optimized expert vs previous evaluator | 24 / 9 / 19 / 8 | 54.8% |
| Hybrid vs optimized expert (800 ms) | 5 / 3 / 2 / 2 | 65.0% |
| Hybrid vs previous evaluator (800 ms) | 8 / 3 / 1 / 0 | 79.2% |
The 120 ms resolved-game score was 58.3% against the optimized expert (paired 95% interval 51.0–65.8%) and 75.9% against the previous evaluator (69.4–82.2%). These intervals are conditional on games finishing. Without assuming outcomes for unresolved games, the overall scores are bounded at 50.0–64.2% and 68.3–78.3%, respectively.
The edge over the expert is modest, and individual presets remain uncertain. The improvement over the previous mixed-piece evaluator is much clearer. The smaller 800 ms tests support the result but do not establish a human rating or a guarantee against any particular player.
What remains unchanged
The proven pawn network, opening book, exact pawn tables, and pawn-only search path. The queen-capture rules remain in force: a newly promoted queen can be captured on the immediate reply; promotion of the last pawn is not an automatic no-pawns draw.
The native rules matched 5,000 JavaScript positions and 77,201 legal continuations. Further replay checks covered 1,924 benchmark games. Tests also verify unchanged pawn assets and decisions, model portability, score decomposition, and the bounded correction. Interactive browser QA was unavailable in this build environment.
Download the complete experiment report · Original pawn-engine research