Source Authorship Forensics

Coda — human-written, or AI-generated?

A source read of adamtwiss/coda, contrasted with Stockfish — the most consequential case in this series.

Author Adam Twiss (direction) + Claude Code (all code) Language Rust Scope ~58,000 LoC · 2,334 commits in ~3.5 months Eval From-scratch NNUE (HalfKA + threat/x-ray features) Analyzed 2026-07-11 Stockfish similarity Strong →

Verdict · AI-Generated (agentic)

Machine-written, expertly directed — the sophisticated end of AI-generated.

Coda states it outright, and the name says it too — Chess Optimised, Developed Agentically. Its README: "every line of code was written by Claude Code, with direction, testing and review by a human." On the authorship axis this is unambiguous: the code is AI-generated.

But it is the polar opposite of a careless paste. This is a rigorous, months-long, SPRT-tested agentic build — 2,334 commits of experiment-driven engineering, directed by a capable human who has written engines before. Where Luna was AI output dumped unreviewed and Owen a one-shot agent build, Coda is the same underlying idea done with real discipline: it is a genuinely strong, sophisticated engine.

Confidence: Absolute on the AI-generated verdict — it is declared in the README and in the engine's very name.

A landmark point on the spectrum: fully AI-written, yet fully human-directed. Compare the careless Luna, the one-shot Owen, and the AI-assisted engines like Triumviratus. See the index →

1It tells you exactly what it is

No forensics required for the headline — the project is built on radical transparency about its method.

"Coda is a strong UCI chess engine, developed entirely through human-AI collaboration: every line of code was written by Claude Code, with direction, testing and review by a human. … I had written a few hobby engines in the past, and I wanted to see how far I could get with a new engine with the help of Claude."

It began in January 2026 as GoChess (in Go, built the same way), then was rewritten in Rust in March. The pace is the tell of the method: ~2,334 commits in about 3½ months — roughly 20 a day — sustained at an elite level of quality.

2The rigour is the story

What separates Coda from every other AI-generated engine here is the discipline of its development loop.

#	Signal	What it shows	Weight
1	README: "every line of code was written by Claude Code"	Declared AI-generated authorship	Decisive
2	2,334 commits in ~3.5 months; a huge log of `experiments:` commits with SPRT results, CI rules, and gap attribution	A rigorous, measured, human-directed testing loop — not a dump	Human-led
3	Directed by a developer with prior engine experience; a predecessor (GoChess) rebuilt in Rust	Real human judgement steering the agent	Human-led
4	A contribution from Andrew Grant (author of Ethereal / OpenBench)	Engaged with the real computer-chess community	Support
5	No Stockfish copyright / fork markers anywhere	Not a derivative — independent (AI-written) code	Support

The "done right" case. If Luna is the cautionary tale of AI output shipped unread, Coda is its mirror image: AI writes every line, but a skilled human runs a real experimental program around it — SPRT-gated changes, falsified hypotheses recorded, regressions rolled back. It is the strongest evidence in this whole series that agentic development, done with discipline, can produce a serious engine.

3And the engine is genuinely sophisticated

This is not a toy. The feature set is at the level of a strong modern engine, with real originality.

Novel NNUE: a from-scratch network combining HalfKA piece-square features (16 king buckets) with ~67k explicit threat and x-ray attack features — "the network sees not just where pieces stand but what they attack" — a 1024-wide accumulator, int8 hidden layers, material-bucketed output heads, trained with a customised Bullet trainer on the LC0 dataset plus self-play. Fully incremental inference with AVX2/AVX-512-VNNI/NEON kernels.
Full modern search: PVS, NMP with verification, reverse futility, razoring, futility/late-move pruning, SEE pruning, ProbCut, internal iterative reductions, threat-shaped LMR, singular extensions (double & negative), cuckoo-table repetition detection.
Rich history: threat-aware 4D main history, capture / four-ply continuation / pawn history, and multi-source correction history.
Systems care: Lazy SMP over a lockless, XOR-verified, huge-page-backed TT with acquire/release atomics — correct on ARM as well as x86 (Apple Silicon first-class). Syzygy + Polyglot; self-play datagen in SF binpack format.

4Compare & contrast with Stockfish

Where it resembles Stockfish
The full modern search stack (NMP, RFP, LMR, SEE, ProbCut, singular extensions)
NNUE evaluation; Lazy SMP; lockless XOR-verified transposition table
Correction history, cuckoo repetition, Syzygy tablebases
SPRT/OpenBench-style, Elo-measured development

Where it is its own (AI-written) work
Original NNUE with explicit threat / x-ray attack features (not a standard HalfKA net)
Written in Rust, from scratch — no Stockfish code (0 copyright markers)
ARM-first atomic correctness; huge-page 5-slot TT buckets
Authored entirely by an AI agent under human direction

Bottom line

Coda is AI-generated — every line written by Claude Code, as its README and its very name declare — but it is the defining example of that done well: an expert human ran a disciplined, SPRT-tested, months-long experimental program, and the agent produced a genuinely strong, original engine (novel threat-aware NNUE and all). On the authorship axis it sits with Luna and Owen in the "AI-generated" band; on every axis of quality and rigour it sits at the opposite extreme. It is the clearest sign yet of where agentic engine development is heading — and shares Stockfish's techniques strongly while borrowing none of its code.