Dorky 5.1 - new version chess engine

Dorky - Dorky 5.0 adds native UCI protocol support and Syzygy endgame

tablebases, and closes the crash and time-forfeit problems present in 4.8.

Author - Matt McKnight  Rating CEDR=2854

What's new?

Dorky 5.1 is about +105 Elo stronger than Dorky 5.0.

Measured directly on the shipping binary against tagged v5.0, matched UCI
protocol, one core, no engine book, UHO_4060_v4.epd repeated-color openings,
transposition hash 256 MB and pawn hash 64 MB:

time control	games	W-L-D	score	Elo	95% CI
10+0.1	4,500	2289-983-1228	64.51%	+103.8	[+95.0, +112.8]
60+0.6	927	454-168-305	65.43%	+110.8	[+92.3, +130.0]
pooled	5,427	2743-1151-1533	64.67%	+105.0	[+97.0, +113.1]

Elo intervals use the trinomial W/L/D standard error and the logistic Elo
transform. This is a single direct measurement of the complete 5.0 → 5.1 span,
not a figure composed from the individual changes listed below.

What changed

The dominant change is the evaluation rewrite.

Evaluation

• Full tapered middlegame/endgame evaluation. Dorky now carries independently
  tuned MG and EG values for the whole evaluator and interpolates them
  continuously over a conventional 24-step material phase. This replaces the old
  partial, term-by-term phase handling. It brought wider ordinary piece-square
  tables, a tapered king table, and a simplified tapered pawn-chain model built on
  color-specific base attacks. The full tune fitted K=1.0211 and cut holdout MSE
  from 0.094405 to 0.093281.
• Outpost redesign, retuned, shipped as a bundle together with two new trap
  terms: Greek-gift sacrifice evaluation and the fishing-pole / trojan
  pattern. The Stonewall structure handling was repaired in the same work.
• Mobility reworked and fully tuned, including a minor-supporting-piece term
  and a centralized mobility area. The rook mobility terms were kept alongside the
  existing rook terms rather than replacing them.
• Separate middlegame and endgame queen values, plus
  BISHOP_PLUS_PAWNS_ON_COLOR and a knight-pair redundancy term.
• Development gradient. While a side still has a minor on its home square, the
  positive half of the knight piece-square table is scaled down, so repositioning
  an already-developed knight gains less than bringing a new piece out. Bishops
  are deliberately excluded: their home squares score near zero, so the same
  scaling would penalize the first developing move instead. A queen half of the
  same shape also ships. Both scales spent part of the release span reverted to
  inert and re-entered as fitted values from the full tune, so neither carries a
  standalone strength measurement.
• Bug fixes: BISHOP_XRAY_HEAVY indexed the slider tables with empty occupancy;
  the heavy-piece xray code had a second defect; king-king tropism was
  de-registered from tuning and restored to its origin value.

Search

• Singular extensions.
• Logarithmic LMR reduction schedule — log(depth) x log(moveCount).
• Mate-distance pruning, plus a fix replacing invented near-mate constants
  with MATE_THRESHOLD, which had been corrupting aspiration-window searches.
• The root aspiration ladder is now a function of the iteration rather than of
  whichever move happens to sit in root_moves[0].

Speed and correctness

• BMI2 PEXT sliding-piece attacks, behind a compile guard, with the rotated
  bitboards retained as a validation oracle. Worth about −5.19% search time.
• UCI protocol performance fix — Dorky was doing per-move work on every move
  under UCI. currmove output is now optional and off by default.
• An SMP nodes_searched race is fixed by counting into a local move count.
• Faster string parsing on the protocol path.
• Castling out of check is no longer generated. The castling legality test
  checked the squares the king crosses and lands on but never the square it
  starts on. Quiescence's checking-move generator could therefore produce a
  castle that escaped check and delivered one simultaneously — a move that does
  not exist — and search it. It never reached the board: the played move comes
  from legal root generation, and the transposition table's move is validated
  before use, so this cost search accuracy rather than legality. The defect had
  been present since the project's first commit in 2015. Measured over 2,584
  games at two time controls, correcting it produced no detectable change in
  playing strength (+1.4 at 10+0.1, +1.3 at 60+0.6); it ships as a correctness
  fix.

Component results are history, not arithmetic

The individually measured ships were LMR +9.0, MG/EG queen value +11.0, full
mobility +6.4, singular extension +9.1, the outpost/trap bundle +14.6, and the
full phase split +44.4. Those point estimates are development history. They are
not a composable rating claim and they do not sum to the release figure. The
direct 5.1-vs-5.0 arm above is the measurement that counts.

Compatibility notes

• Dorky 5.1 requires a BMI2-capable CPU — Intel Haswell (2013) or newer, AMD
  Excavator (2015) or newer. The PEXT slider index has no runtime fallback, so
  older hardware that ran 5.0 will fail on 5.1 with an illegal-instruction crash.
  AMD Zen 1 and Zen 2 run correctly but microcode the instruction and gain
  nothing from it.
• The Linux archive ships no opening book. book create does not work on the
  Linux build, and a book is not portable between platforms — the footer stores
  its cluster length in a C long, 4 bytes on Windows and 8 on Linux, so each
  rejects the other's book.bin. The Linux engine plays from move one with no
  error; GUI-supplied books work normally. The Windows archive ships a book as
  before.
• UCI currmove is no longer emitted unless enabled. GUIs that displayed the
  current root move will show nothing there by default.
• Dorky.ini defaults remain transposition_table=256, pawn_table=64. Every
  number in this document was measured at those sizes.

Verification

• Strength: the arm above, 5,427 games across two time controls.
• Determinism and tactics: four fixed-depth suites reproduce their expected
  node signatures exactly on the release binary — games.epd d10 483,965,181,
  wac.epd d10 78,125,927, and the 100-position depth subset 717,554,936 at
  depth 14 and 2,892,203,018 at depth 16. All three recorded perfts return their
  exact values: start position depth 6 = 119,060,324, Kiwipete depth 4 =
  4,085,603, promotion-dense depth 4 = 422,333.
• Tablebases: scripts/check-tablebases.ps1 passes 1,191 of 1,191
  positions, with Syzygy and Nalimov agreeing on every shared position, the
  hand-annotated anchors holding, and 5-man Syzygy-only fallback behaving.
• Trap evaluation: trap-eval.epd with scripts/check-trap-eval.ps1 compares
  a live build against a compile-time detector stub. Positive controls, mirrored
  by color and wing, cover fishing-pole SET and SPRUNG, Greek-gift BASE,
  SACKED, and CONTROL; one-condition-negative controls cover each required
  precondition individually. Thirty live-versus-stub root-move comparisons at
  depths 8 through 16 were identical, so neither detector induces a sacrificial
  root choice. Temporary instrumentation over games.epd recorded 132,531
  fishing-pole and 71,487 Greek-gift scoring events, confirming both detectors
  fire in ordinary play; that instrument does not ship.
• Runtime: a multi-core Release VALIDATE soak across UCI and XBoard, ponder
  on and off, and with a Syzygy path configured.

Known limitations and deferred work

• Drawish material signatures are not scaled. Configurations such as rook
  versus minor are still scored on their raw material difference rather than
  pulled toward zero.
• Won pawnless endgames convert slowly. This is a separate, open problem.
• Also deferred past this release: the blocked-pawn redesign, aspiration-window
  experiments, further outpost tuning, and NNUE.

Dorky 5.1 download   mirror