Source Authorship Forensics
A source-level read of dylan2554/onyx, contrasted with Stockfish.
Onyx is refreshingly honest about its own authorship: the engine’s UCI identity literally reports id author Dylan (with Claude). It is a competent single-file C++ engine (magic bitboards, PVS, LMR, PeSTO evaluation) fed by a network the author trained himself, and it descends from the author’s own earlier “Sable” engine (the NNUE format still uses SABLE1/2 magic bytes).
But it lands as finished files via the GitHub web UI — 21 “Add files via upload”/“Delete” commits, no incremental history — so the code was substantially produced with an AI assistant rather than hand-developed in the open. Real human direction (own net, own lineage, open disclosure); heavy AI co-authorship.
Confidence: Medium-High — the self-disclosure is explicit; the exact human/AI split is not.
No forensics needed for the headline — Onyx says so in its own UCI banner.
// onyx.cpp
printf("id author Dylan (with Claude)\n");
| # | Signal | What it shows | Weight |
|---|---|---|---|
| 1 | UCI id string: author Dylan (with Claude) | Self-disclosed AI collaboration | Nuance |
| 2 | Own self-trained NNUE (i7-11700K + RTX 3070); descends from the author’s earlier “Sable” engine | Genuine human ownership and lineage | Strong |
| 3 | All 21 commits are GitHub web-UI uploads/deletes; single monolithic onyx.cpp | No incremental history — code arrives finished | Nuance |
| 4 | Competent, conventional engine (magic bitboards, PVS/LMR, PeSTO eval) | Well-formed output — as expected from a good assistant | Support |
| Where it resembles Stockfish |
|---|
| Bitboards + magic move generation |
| PVS / alpha-beta, TT, null-move, LMR, killers/history, quiescence |
| Self-play-trained NNUE evaluation |
| Aspiration windows |
| Where it differs |
|---|
| Single 1,835-line file vs a large modular codebase |
| PeSTO tapered eval as the classical base |
| Substantially AI-co-written vs decades of human work |
| Uploaded as finished files, no open dev history |
Onyx sits in the AI-assisted middle of the spectrum — and says so itself. A real developer trained his own network, carried forward his own engine lineage, and openly credits Claude as co-author. It differs from the expert, incrementally-developed Triumviratus (AI as a tool in a long human loop) by leaning further toward AI co-authorship: competent code, honestly disclosed, delivered as finished files.