CeylonDemon 2.0

CeylonDemon 2.0 is a free UCI chess engine with Resonance, a dual-frame NNUE
evaluator that reads every position through two feature transformers — one
anchored on the side to move's own king, one on the opposing king — and reports
the residual between them as an uncertainty signal the search consumes directly.

Powered by the NARC Engine.

Downloads

File	Requires
CeylonDemon-2.0-x86-64-avx2.exe	x86-64 with AVX2 — portable release build, use this if unsure
CeylonDemon-2.0-x86-64-bmi2.exe	x86-64 with AVX2 + BMI2 — Skylake and later, fast PEXT

The neural network is linked into both executables, so each is a single
self-contained file — there is nothing else to download and no network file to
configure.

Strength

2719.5 ± 32.8 — a locally transferred CCI-STC estimate under 10+0.1
conditions.

This is not an official CCI rating and not universal Elo. It is an
interpolation against a fixed set of anchor engines.

Games	320, all completed — no stalls, crashes or forfeits
Record	39–82–199 (80.0 / 320, 25.0%)
Conditions	10+0.1 s, 1 thread, 64 MB hash, UHO openings, no adjudication
Opponents	8 anchors from 2761 (Monolith 3.0) to 3101 (Rose 1.0.0)
Tools	FastChess 1.8.2-alpha, Ordo 1.2.6 (-F 95, 10,000 simulations)
Hardware	Intel Core i7-6600U @ 2.60 GHz, 2C/4T

The engine sits inside the anchor range rather than below it — outscoring
Monolith 3.0 at 43.8% and losing to everything above — so the rating is an
interpolation rather than an extrapolation. Seven of the eight per-opponent
residuals fall between −0.6 and +3.5 points. The exception is Rose 1.0.0, against
which 2.0 scored 0.5/40 where its rating predicts about 4.0/40; that is a genuine
outlier outside sampling variation and the cause is not yet identified.

Caveats. The ±32.8 is statistical error only; the error introduced by
transferring against these particular anchors is systematic and unquantified.
The result says nothing about other time controls, thread counts or hash sizes,
and version gaps under roughly 65 Elo are unresolvable at this precision.
CeylonDemon has never played a game under CCRL conditions and does not appear on
any public rating list. Independent testing is welcome.

Which build was measured. The rating run used the AVX2 build. The BMI2 build
is compiled from the same source and differs only in instruction selection; it
has not itself been through a rating run. Both builds were verified to produce
identical evaluations and identical best moves on the positions checked below.

Verification

CeylonDemon-2.0-x86-64-avx2.exe
  SHA-256  ab7fb2eab6012c380873a1b416ba499b3cbe7172635b539dc615b8160fa3d979
  Size     13,184,680 bytes

CeylonDemon-2.0-x86-64-bmi2.exe
  SHA-256  37912ebb5844ced69c3711bb3959801b53b08ecf09717ab3f4e8eb6e66d3033b
  Size     13,183,656 bytes

Both binaries are compiled from the source at this tag with GCC 15.1.0
(MSYS2 UCRT64), statically linked. Move generation was verified against the
published perft references before release:

Position	Depth	Nodes
Start position	5	4,865,609
"Kiwipete"	4	4,085,603

The AVX2 binary differs from the specific executable used in the rating run in
4 bytes only — the PE build timestamp and the resulting header checksum — with
identical code, data and embedded network.

Engine details

• UCI options: Hash (128), Threads (1), Move Overhead (30), EvalFile
  (<embedded>)
• Resonance v10 evaluator, integer inference throughout, embedded network
• Lazy SMP, 1–256 threads

Provenance

The embedded weights were warm-started from a Stockfish teacher network. The
engine source contains no Stockfish code, but the network carries Stockfish
ancestry and is not presented as independently derived. A clean-provenance
network is under development. See NOTICE.md for full attribution, including
the search techniques the engine implements but did not originate, and the
disclosure of AI-assisted development.

Licence

GNU General Public License v3 or later. Copyright © 2026 Madushan Dissanayake.