ArhanPassant

Engine · version 0.13.0-rl4

Built to be measured

ArhanPassant is a UCI chess engine written in Rust. Every claim about its strength comes from games it actually played, and every new version has to beat the previous one before it ships.

Move generation

Legal moves from magic bitboards, matching 21 published perft results, up to the 706,045,033-position test.

Search

Iterative deepening with aspiration windows, principal variation search, quiescence search, a shared lock-free hash table and lazy SMP threads.

Pruning and ordering

Null-move, reverse futility, futility, late-move and static-exchange pruning; late-move reductions; singular extensions; history, continuation-history, threat-aware history, killer and counter-move ordering.

Evaluation

Champion 0.13.0-rl4 uses a 512-hidden-unit NNUE network with eight king buckets and eight output buckets. It grew from 0.12, trained on 381.8 million self-play positions, through four rounds of reinforcement learning on fresh self-play. Every new network must pass a strength gate before promotion.

Interfaces

UCI for chess GUIs, a Rust library, and champion 0.13.0-rl4 in this site's WebAssembly engine, with SIMD and a portable fallback.

Measured strength

Rating estimate · 7 October 2026

≈ 3,507

on the CCRL Blitz scale, measured in our own matches

95% interval 3,472 to 3,541 · 900 games at 10+0.1 seconds

Measured build: network 0.13.0-rl3 with three search changes that passed their own tests and are now being checked together.

Latest promotion · v0.13.0-rl4

+6.7 Elo

Versus v0.13.0-rl3, with the same engine code and search settings, over 5,714 games at 5+0.05 seconds. The 95% interval is +1.5 to +11.9 Elo. The network passed its sequential strength gate.

How the estimate is made

ArhanPassant played 100 games against each of nine engines at 10+0.1 seconds, every engine on one thread with 64 MB of hash, from the 0.12 benchmark's opening positions with colours swapped. Each opponent it scored between 20% and 80% against implies a rating from that engine's published CCRL Blitz rating, and the implied ratings are combined, weighted by their precision. This run used Koivisto 9.0, Ethereal 12.00, Stash 34.0, Stash 37.0 and Weiss 2.0. The interval covers game statistics only.

This is our own measurement, not a CCRL rating. CCRL has not tested ArhanPassant, and its testers use other hardware and time controls.

Same method, over time

DateBuildEstimate
5 October 20260.12.03,456 (3,433–3,479)
6 October 20260.13.0-rl13,487 (3,453–3,520)
7 October 20260.13.0-rl3 + search changes3,507 (3,472–3,541)

0.12 played 200 games per opponent, later builds 100. Which opponents fall inside 20–80% changes as the engine improves, so the like-for-like comparison is the table below.

Against the same engines, before and after

Champion 0.12 played 200 games against each of these engines on 5 October 2026. The measured build played 100 against each on 7 October 2026, with the same openings, time control and settings. Averaged over all nine opponents, its results are +70.6 Elo better (95% interval +46.7 to +94.5).

Score is ArhanPassant's share of the points. A single row rests on 100 to 200 games, so its interval is wide; the average is the steadier number. Stockfish 19 is not on the CCRL list we use, so it counts in the comparison but not in the rating.

OpponentCCRL Blitz0.12 scoreNow: won · lost · drawnNow scoreChange, Elo (95%)
Stockfish 19—12.0%1 · 73 · 2614.0%+30.8 (−36.2 to +97.7)
Stockfish 17.13,77111.8%0 · 66 · 3417.0%+74.8 (+16.5 to +133.1)
Koivisto 9.03,63216.3%9 · 57 · 3426.0%+103.1 (+34.4 to +171.8)
Ethereal 12.003,39065.0%61 · 12 · 2774.5%+78.7 (+14.9 to +142.6)
Laser 1.73,29272.5%75 · 9 · 1683.0%+107.0 (+27.2 to +186.9)
Stash 34.03,31568.3%66 · 13 · 2176.5%+72.1 (−3.0 to +147.2)
Stash 37.03,41947.5%46 · 30 · 2458.0%+73.5 (+15.5 to +131.5)
Weiss 1.43,25580.8%79 · 9 · 1285.0%+52.2 (−43.8 to +148.3)
Weiss 2.03,31871.8%64 · 11 · 2576.5%+43.1 (−26.2 to +112.4)

Historical cloud rating estimate

Earlier champion 0.9 configuration, against engines with published CCRL Blitz ratings, at 60+0.6 seconds

≈ 3,426

95% interval 3,414 to 3,437 from 2,284 historical games

Each opponent's rating is read from the CCRL Blitz list, and one rating is fitted to the matches the engine scores between 20% and 80% in. At 10+0.1 seconds, 5,200 games give 3,448 (3,436 to 3,459). These are internal estimates from cloud matches. CCRL has not independently rated ArhanPassant; this older result does not rate the current version.

Opponents scored below 20% against in the earlier fast run: Stockfish 17.1 (CCRL 3,771) 3.9%, Koivisto 9.0 (CCRL 3,632) 10.6%. An earlier measurement of version 0.7.0 against Stockfish 19's calibrated levels gave about 3,026.

OpponentCCRL BlitzWon · lost · drawnScore
Stash 37.03,41972 · 104 · 21245.9%
Ethereal 12.003,390143 · 78 · 16758.4%
Weiss 2.03,318170 · 57 · 16164.6%
Stash 34.03,315108 · 18 · 11868.4%
Laser 1.73,292122 · 26 · 9669.7%
Weiss 1.43,25579 · 11 · 5473.6%
Demolito 20213,243186 · 9 · 4986.3%
Stash 31.03,208177 · 16 · 5183.0%

The strength gate

A candidate version plays the current champion in pairs of games: same opening, colours swapped, from the UHO_4060_v4 opening book at 5+0.05 seconds. After every pair a sequential probability ratio test compares two hypotheses: the candidate is no better (0 Elo) or it is at least 5 Elo stronger.

The test stops as soon as the evidence is strong enough either way. With α = β = 0.05, a version that is no better gets through about 5% of the time, and a version that really is 5 Elo stronger passes about 95% of the time.

Versions up to 0.12.0 were tested the same way at 8+0.08 seconds.

Checked before use

  • On 600 simulated matches with known strength: 5.0% false passes, 95.2% true passes, 50.7% at the midpoint.
  • The engine against an identical copy: exactly 0 Elo over 300 games.
  • 4× the search budget against 1×: passed after 414 games; the reverse failed after 408.

Every search change, tested

Changes to the search and native evaluation go through the same test: the engine with the change plays the engine without it, with the same network and equal resources. Accepted, rejected and undecided tests are listed here. Correctness checks and small scouting matches help select experiments; they do not prove a strength gain.

ChangeGamesElo (95% interval)Result
Time management: scale the soft limit by the best move's share of the search and by how long it has stayed best1,888+15.5 (+7.0 to +24.0)Passed
Evaluation correction: learn, per pawn structure, how far search results land from the static evaluation, and correct for it960+29.4 (+17.1 to +41.7)Passed
ProbCut: at depth 5 and above, a capture that beats beta by 200 at reduced depth cuts the node3,872−4.0 (−10.1 to +2.0)Rejected
Razoring: at depth 4 or less, a position 250 per ply below alpha is checked with captures only3,840−3.6 (−9.8 to +2.5)Rejected
Evaluation correction also learned per piece set of each side (pieces other than pawns)4,736+7.9 (+2.3 to +13.5)Passed
Evaluation correction also learned per previous move448−70.8 (−89.3 to −52.6)Rejected
After a reduced search fails high, re-search one ply deeper or shallower depending on the margin12,000+4.4 (+1.0 to +7.8)Passed
Quiescence futility: skip captures that cannot reach alpha even with a 150 margin896−25.6 (−38.5 to −12.8)Rejected
Mop-up: a bare king against mating material is driven to the edge (bishop and knight: the bishop's corner); converted 7 of 7 such endings at 100 ms a move, against 2 of 7 without6,240+0.3 (−3.8 to +4.5)Kept: no loss
A table cutoff by a quiet move rewards that move's history, as a search cutoff would12,128+1.0 (−2.3 to +4.4)Undecided
A node that fails low rewards the opponent's quiet move that led to it7,904+5.2 (+1.0 to +9.4)Passed
The opponent's quiet move learns from the evaluation swing it caused3,296−4.2 (−10.7 to +2.3)Rejected
Quiet moves are reduced one more ply when the table's move is a capture12,096+2.8 (−0.6 to +6.3)Undecided
Exact NEON neural output on Apple Silicon, using i64 reduction; compare against the saved scalar build at equal time control and network448+80.5 (+63.2 to +98.3)Passed
Continuation history also keyed by our own move two turns ago7,296−0.6 (−5.0 to +3.8)Rejected
History pruning: at low depth, skip quiet moves whose history keeps failing12,000+3.0 (−0.6 to +6.6)Undecided
Eval correction tables learn jointly, each toward what the others leave, instead of each the whole error3,200−4.8 (−11.6 to +2.1)Rejected
Eval correction also learned per previous move, with the tables learning jointly3,072+10.2 (+3.1 to +17.2)Passed
Speed: faster legality checks, move lists left uninitialised, a 16-bit network output kernel for NEON and AVX2, and hash-table prefetch; same search, more positions per second504+28.3 (+11.4 to +45.4)Passed
Move ordering: quiet-move history kept separately for threatened squares and for each pawn structure; quiet moves that save a piece attacked by a cheaper one, and safe checks, are tried earlier2,676+11.2 (+3.7 to +18.6)Passed
Six search refinements together: fail-high blending, an extra reduction after repeated cutoffs, capture-history reductions, stronger negative singular extensions, history-aware pruning depth, and an 'improving' fallback1,220+2.3 (−8.6 to +13.2)Undecided
Three of those refinements: fail-high blending, an extra reduction after repeated cutoffs, and the 'improving' fallback720±0.0 (−14.0 to +14.0)Undecided
Time management: think longer when the score falls1,080−3.5 (−14.7 to +7.6)Undecided
Network with 16 king buckets instead of 8, trained on the same positions2,002−9.2 (−17.9 to −0.5)Rejected
Quiescence search: after three moves, skip captures that are not recaptures, checks or promotions4,174−3.2 (−9.3 to +2.9)Rejected
Evaluation correction keyed by the knights and bishops on the board6,512+6.1 (+1.3 to +11.0)Passed
A side that can force a repetition scores at least a draw, detected ahead of time with cuckoo tables2,640+11.2 (+3.7 to +18.7)Passed
Capture futility pruning near the leaves, base margin 2501,254+19.1 (+8.8 to +29.5)Passed
With four threads, helper threads stagger their search depths2,060+1.5 (−6.3 to +9.4)Undecided
Search margins and reductions from 6,000 pairs of automatic tuning (SPSA), tested together8,000+4.7 (+0.2 to +9.2)Undecided
Pruning: keep quiet moves that give check even when the static exchange says they lose material6,648−1.1 (−5.9 to +3.7)Rejected

How it improves now

Reinforcement learning from self-play

The champion plays itself at 8,000 nodes per move on Oracle Cloud virtual machines and a Mac. Once 15 million new positions are in, the champion network is fine-tuned on them plus the previous round's, and the result must pass the strength gate before it replaces the champion. Gains so far, round by round: +27.3, +20.5, +16.0, +6.7 Elo.

Learning from its losses

Two of the Mac's eight self-play threads start their games from 92,906 positions taken from games ArhanPassant lost in its benchmark matches, so training sees more of the positions it handles worst.

Search changes

Each search idea ships switched off and becomes a default only after its test passes. The tests above list what was kept, what was rejected and what is still undecided.

Parameter tuning

SPSA tunes the search parameters on spare cores. Tuned values have to pass the same strength gate before they are used.

Lichess play is paused while engine development and testing continue.

Use it

In a chess GUI

Download version 0.13.0 for Windows, Linux or macOS, or install it with Homebrew or Docker, then add it as a UCI engine in Cute Chess, Arena, Banksia or En Croissant.

brew install arhancanli/tap/arhanpassant
docker run -i --rm ghcr.io/arhancanli/arhanpassant

# or build it yourself
git clone https://github.com/arhancanli/arhanpassant
cd arhanpassant && git checkout v0.13.0
cargo build --release
./target/release/arhanpassant bench

UCI options

Hash16Hash table size in MB (1 to 1,048,576).
Threads1Search threads (1 to 512).
Move Overhead30Milliseconds kept in reserve per move for lag.
EvalFile<embedded>Path to a network file, or <none> for the hand-written evaluation.
SyzygyPath<empty>Folders of Syzygy tablebase files, separated by colons.
UCI_Chess960falseChess960: castling as king-takes-rook, X-FEN and Shredder-FEN.