Source Authorship Forensics
A source-level read of HSR-projects/Owen v1.0.0, contrasted with Stockfish — the fourth in a series.
The repository carries a literal fingerprint of an AI coding agent: the C++ engine's package.json declares a dependency on the claude npm package, with node_modules/claude committed into the tree. A C++ engine has no legitimate use for that — it is residue of the project being built with the Claude Code agent. Taken with a single 11,472-line initial commit and a placeholder author identity, the authorship is agent-generated.
Unlike a careless dump, this is a capable, well-directed build: it integrates real third-party components correctly, documents what is original versus borrowed, and ships a working engine. A human clearly steered it — but the code was produced by an AI agent, not hand-written.
Confidence: High — the agent artifact plus the single-dump history are mutually reinforcing.
The clearest evidence isn't in the engine code — it's in the build metadata.
// engine/package.json (in a C++ project!)
{ "dependencies": { "claude": "^0.1.1" } }
// engine/node_modules/claude/package.json — committed into the repo
{ "name": "claude", "version": "0.1.1",
"repository": { "url": "git+https://github.com/bcherny/redirect-claude.git" } }
claude. Its presence — plus the committed node_modules/claude — means the working directory was an AI-agent (Claude Code) workspace, and the artifact was committed along with the generated engine.| # | Signal | What it shows | Weight |
|---|---|---|---|
| 1 | Committed claude npm dependency + node_modules/claude | Repo built inside a Claude Code agent workspace | Smoking gun |
| 2 | Entire engine in one commit — 54 files, 11,472 insertions | No incremental human development history | Strong |
| 3 | 3 total commits, all one day; author email hemesh@example.com | Placeholder identity; no real dev timeline | Strong |
| 4 | Uniform polish; comprehensive, marketing-style README | Consistent with generated-then-documented output | Support |
Owen is not Luna. The output is competent and honest about its parts.
The README carefully separates original work from third-party: the board, move generation, and search are presented as Owen's own; the evaluation is explicitly Stockfish's HalfKP NNUE via the nnue-probe library, and tablebase probing is the Fathom library — both included with attribution. The C++20 core is clean and idiomatic (namespaced, constexpr, tightly-packed TT entries, Lazy SMP with a shared table and atomic stop flag), move generation is perft-validated by the test suite, and a separate from-scratch AlphaZero-style trainer lives in training/. This is what a well-directed agent build looks like — the opposite end of the quality range from an unreviewed paste.
Owen doesn't just resemble Stockfish — in part it runs on Stockfish.
| Shared / borrowed from Stockfish |
|---|
| Uses Stockfish's HalfKP NNUE network directly for evaluation (via nnue-probe) |
Stockfish-style engine idioms: COLOR_NB/SQUARE_NB, TT flags, killers/history |
| Modern search stack: PVS, NMP, LMR, aspiration, SEE, futility, check extensions |
| Lazy SMP with a shared transposition table |
| How it differs |
|---|
| Authored by an AI agent in a single pass vs decades of human contributors |
| ~9k LoC, one release vs a vast, continuously-tested codebase |
| Relies on Stockfish's net rather than training a competitive one of its own |
| Syzygy via the Fathom library rather than a bespoke implementation |
Owen is an AI-agent-generated engine — built with Claude Code, as the committed claude dependency and one-shot 11.5k-line history make plain — but a competent, human-directed one that integrates real components honestly and runs. On the spectrum from Luna's careless AI paste to the hand-built human engines, Owen occupies a distinct spot: machine-authored, yet well-supervised. Its evaluation strength is, quite literally, borrowed from Stockfish.