Source Authorship Forensics

Chess Engine Forensics — about the project

How the collection is made: the two axes, the methodology, the coverage, and where everything is saved.

A companion project on this site: a source-level read of chess-engine repositories, judging each on two independent axes — authorship (human, AI, or in between) and Stockfish similarity (independent vs. built on Stockfish). Every engine is cloned, its git history and code scanned for tell-tale fingerprints, and contrasted with Stockfish. Live index: /engine-analysis/.

The two axes

Authorship spectrum

Band	Meaning
AI-generated	The code itself was produced by AI; the human role is direction / prompting.
AI-assisted	A person's own project, with an AI assistant used as a tool in the loop.
Human-written	Authored by a person — hobbyist to expert to veteran.

Stockfish-similarity scale

Rating	Meaning
Derived	Builds directly on Stockfish's code or network.
Strong	Re-implements Stockfish's techniques independently (bitboards + NNUE + LMR/NMP/SEE).
Moderate	Modern architecture with its own distinct choices.
Independent	Own board representation / hand-crafted eval / unusual language.

The axes are orthogonal: a human engine can be very Stockfish-like (Catalyst), and "Strong similarity" means shared technique, never shared code.

Methodology — the "tests"

For each engine:

Clone the repository (shallow for large ones).
Git history — commit count, span, authors, message style. Incremental SPRT/OpenBench-style patches over months ⇒ human; a single wholesale dump or all-web-UI uploads ⇒ suspect.
AI-marker scan — grep the source for generated-text residue:
- Gemini [cite: N] markers (Luna had 147).
- A committed claude npm dependency / node_modules/claude (Owen).
- A CLAUDE.md project file (Triumviratus, Arasan).
- Self-disclosure in the UCI banner (Onyx: id author Dylan (with Claude)).
Code texture — jokes, typos, abandoned experiments, tuning notes (human) vs uniform, over-explained, plausible-but-shallow output (AI).
Engineering depth — are the hard parts actually right (e.g. an incremental NNUE accumulator) or merely imitated in shape (Luna recomputed from scratch)?
Stockfish-derivation scan — grep source for Stockfish's copyright header / "part of Stockfish" notice / copied NNUE code. Zero open-source engines examined are forks.

Verdicts infer authorship from convergent evidence; they are well-supported readings, not certifications.

Coverage

New-release engines — full individual write-ups

Mostly from the TalkChess "New engine releases H1 2026" thread.

Engine	Authorship	SF sim.
Luna	🔴 AI-generated (careless paste, Gemini; dev-confirmed)	Moderate
Owen	🔴 AI-generated (agent-built, Claude Code)	Derived
Triumviratus	🟠 AI-assisted (expert human-led)	Derived
Onyx	🟠 AI-assisted (self-disclosed)	Moderate
Askaig	🟠 AI-assisted (self-disclosed, Claude Fable 5)	Strong
Pawnstar	🟠 AI-assisted (veteran-led, since 2014)	Strong
Amira	🟠 AI-assisted (human + LLM)	Independent
rudim	🟢 Human	Strong
Crustik	🟢 Human (hobbyist C)	Independent
Catalyst	🟢 Human (expert)	Strong
tomitankChess	🟢 Human (9-yr JS lineage)	Independent
Ember	🟢 Human (collaborative)	Strong
Lozza	🟢 Human (veteran JS)	Moderate
chessnix	🟢 Human (Zig)	Independent
Ratsu	🟢 Human (Rust)	Independent
Gyatso	🟢 Human (Nim)	Moderate
Porcupine	🟢 Human (Rust)	Independent

CCRL 40/40 top 50 — survey + standouts

Survey: /engine-analysis/ccrl-top50/ — 40 of the top 50 confirmed open-source, cloned & scanned (8 commercial/closed, 2 with no findable source; 40 + 8 + 2 = 50).
Standout pages: Stockfish (the reference), Arasan (human since 1994 but ships a CLAUDE.md — the only AI-marker hit in the whole elite sweep), Chess System Tal (this site's own engine, CCRL #19).

Headline findings (elite tier): 0 AI-generated engines (no model autonomously builds a 3500-Elo engine); 1 AI-assisted (Arasan); 0 actual Stockfish forks — the top has converged on the paradigm, not copied the code. Language variety: mostly C++, strong Rust presence, and Lizard is C# (top 12).

Site structure

Each analysis is a self-contained HTML page (inline CSS, theme-aware light/dark, responsive, GoatCounter view-counter) at /<slug>-analysis/.

/engine-analysis/ — the overview index (authorship spectrum, verdict table, Stockfish-similarity spectrum, methodology).
/engine-analysis/ccrl-top50/ — the CCRL survey table.
Homepage "Chess engine forensics" section (#engine-forensics) with a summary blurb and three cards, plus a nav link.

Page template & tooling

The pages share one CSS/scaffold. Generators are archived in tools/ in the repository: gen_pages.py (the TalkChess pages), gen_ccrl.py (the CCRL standouts), and add_sfsim.py (injects the "Stockfish similarity" header line into each page).

To add a new engine: clone & run the tests → pick a verdict + SF-similarity → author the page (copy an existing *-analysis/ as a template) → add a row/link to /engine-analysis/ (and the CCRL survey if applicable) → deploy.