Repository navigation
Cadence 0.6.0, built from 170f5e3 at the tag 0.6.0. Bench: 34686410 nodes.
Seven binaries: Windows and Linux x86-64, each in an AVX2 build and a generic one for processors without AVX2 (it still needs POPCNT), Linux aarch64, macOS arm64, and macOS x86-64 in an AVX2 build. SHA256SUMS lists their checksums, with those of bench.txt, which holds the bench count every binary was checked against, and of the install scripts. Each binary carries a build provenance attestation: gh attestation verify <file> --repo grounzero/cadence.
The macOS binaries are not notarized. Clear the quarantine flag with xattr -d com.apple.quarantine <file>, or build from source.
install.sh (Linux, macOS, and Windows under Git Bash or WSL) and install.ps1 (Windows PowerShell) install the binary for the system they run on once its SHA-256 matches SHA256SUMS and its bench matches bench.txt. Where no binary fits or the check fails, as on an Intel Mac without AVX2, they build the release from source and hold the build to the same bench. Without --version they take the latest stable release, or while none is published they offer the newest pre-release; --version 0.6.0 names this one. Their SHA-256s, as in SHA256SUMS:
c9fc68954fd1eef3906661002bfe3c841fea3b3f44474ccc3975c835094634c0 install.sh
d2f3aceb7a501f9ad326146bf31f12fafbe4fbf627d4a05cffb946b9c001e6e0 install.ps1
From source: git clone https://github.com/grounzero/cadence.git, git checkout 0.6.0, cargo build --release; the binary is target/release/cadence. On an x86-64 processor without AVX2, build with cargo build --release --config .cargo/x86-64-generic.toml.
Everything since 0.5.23, published first as four release candidates, each
a pre-release: 0.6.0-rc.1 (2026-11-06), 0.6.0-rc.2 (2026-11-08),
0.6.0-rc.3 (2026-11-10) and 0.6.0-rc.4 (2026-11-13). It adds release
binaries and install scripts, scores a position where neither side has the
material to mate as a draw, and fixes how the engine answers the GUI, uses
the tablebases and ponders.
Added
- Tests: proofs, checked by the Kani model checker on every change, that
the score encoding, the transposition table's entries and buckets, the
history and correction updates, the pruning margins and reductions, and
the time budget stay within their bounds. A failed proof fails the build. - The README installs from the latest release's own scripts.
- The installers verify each binary (checksum, bench and, with gh, its build
attestation) and say which checks ran. - They upgrade older installs, report when you're up to date, and ask before
a downgrade or replacing a different build. - Clearer errors when a system blocks the binary, and uninstall removes only
what the installer added. - Releases: each release builds seven binaries, Windows and Linux x86-64
each in an AVX2 build and a generic one, Linux aarch64, macOS arm64 and
macOS x86-64, and publishes them withSHA256SUMS,bench.txt, the two
install scripts and a provenance attestation for each binary, made by the
release workflow run at the release's tag; a manual run drafts and attests
a release only when run at its tag, and never once it is published. - The test log:
docs/testing/sprt-log.csvgives every SPRT, passed or
failed, with its games and result.
Changed
- Search: a position where neither side has the material to mate is scored
as a draw. The bench rises from 24,819,899 to 34,686,410 nodes. - Tools:
setup-worker.pyrefuses a configuration inside the repository,
and git ignoresworker.jsonandcredentials.cadence. The Makefile
clears an inheritedRUSTFLAGS, and with itCARGO_ENCODED_RUSTFLAGS
and everyCARGO_TARGET_<triple>_RUSTFLAGS. - The README gives a table of strength estimates from our own gauntlets
against rated engines, anddocs/calibration/strength.mdpublishes each
gauntlet's games and fit, and how to check them. - Tools: the pre-commit hook rebuilds its boundary checker when any file
the checker is built from is newer than it, not only its main source. - The README is shorter. Getting started offers a release's binaries first,
then the install scripts, and the README ends with acknowledgements and
the license. Setting up an OpenBench worker and creating a test are
explained bypython3 openbench/setup-worker.py --helpinstead. - UCI: under
UCI_LimitStrength, the engine prints only the lines the GUI
asked for, and reports the move it plays. - UCI:
SyzygyPathis taken exactly as sent, spaces included. - The README documents every UCI option, the tools and the generic build.
- Search: its state is grouped into its counters, its move-ordering
tables, the run's result and its clock, with no change to play: every
bench position searches the same nodes as before.
Fixed
- UCI:
positionsent during a search stops and joins that search first.
The search could answerbestmovefor the old position after the new
one had been set. - UCI:
setoption name Hashsent during a search stops and joins that
search before the table is resized. The search could end without its
bestmove. - Tools:
perft --dividewrites castling as the position does, the king
taking its rook only where a castling king or rook stands off its
standard square. A standard position listede1h1ande1a1. - Tools:
datagen extractpublishes its training and holdout files as a
pair, so a failure between the two no longer leaves one of them
published. - Tools:
benchruns its searches on the 16 MB stack the engine's search
threads have, as the tests now do, rather than the system's default. - Pondering: after a ponderhit, the iteration the search was running stops
at the move's planned time until an iteration completes. It could run on
to the limit of half the remaining clock and then play the ponder's last
result. - UCI: after a ponderhit, the reported
timeandnpscount from the
go, asnodesdoes.npswas overstated. - Tablebases: at a won root, only the winning moves that reach a capture
or a pawn move soonest are searched. The losing side is unchanged. - Search: when the tablebases cut a node, the transposition table keeps
that node's move. - Tablebases: a root the tables rank reports the tables' verdict as its
score, a root move is ranked under the fifty-move counter as the
tablebase library ranks it, each table file is opened once and one that
fails is named, and stop and the clock are checked between root probes. - UCI: the clock starts when
gois read, and everygois answered,
withbestmove 0000after apositioncommand that failed. A clock
value that is not a whole number reads as zero, agolimit too large
to hold reads as the largest the engine holds, the count ingo mateis
clamped, a root with no legal move reports a score line, and an input
line is limited to a mebibyte. - Positions: a FEN with impossible material is refused, castling letters
with no rook to castle with are dropped, and the move counters are
clamped. - Search: a reduction stays below the depth it reduces, a finished line is
kept, and under a node limit with helper threads the main thread
finishes its first iteration.perftrefuses a depth deeper than the
search allows. - Version: a build calls itself by the release's version only at that
version's annotated tag, on an unmodified tree. - Tools:
bench,perft,spsaandmagicsexit with a failure when
they cannot write their output,texelchecks its arguments and its
weights, anddatagencaps its workers, keeps its games in order and
refuses a record whose moves do not end as it says. - Build: the generic Windows build keeps SSE3.
Test 76, for the transposition table fix: -0.27 (between -2.75 and
+2.20), a self-play estimate against the build before it at the short
control, both sides given the tables up to five pieces. This was a
non-regression test, which asks only whether the fix costs strength. It
was stopped at 24,106 games before it decided, and the fix was taken on a
result that shows no cost. It is not a rating.
Test 77, for keeping the fastest wins at a won tablebase root: +0.36, give
or take 1.63, a self-play estimate against the build before it at the
short control, both sides given the tables up to five pieces. This was a
non-regression test, which asks only whether the change costs strength,
and it passed at 54,626 games. It is not a rating.
Test 80, for scoring a position with no mating material as a draw: +2.73
(between -0.46 and +5.93), a self-play estimate against the build before it
at the short control. This was a non-regression test, which asks only
whether the change costs strength, and it passed at 15,760 games. It is not
a rating.