Source Authorship Forensics
A source-level read of ExxDreamerCode/Ember v1.1.1, contrasted with Stockfish.
Ember is a personal learning project (its Russian-language README says as much) that grew a real collaborator: two distinct developers commit here, exchanging work through actual pull requests. The history is full of quantified, incremental engineering, and the project includes a hand-curated regression suite so distinctive — complete with video demos of the engine's own blunders — that it could only be a person's work.
No AI residue of any kind: no generated-text markers, no agent artifacts, no incoherence. Just two people building an engine and documenting its mistakes carefully.
Confidence: High — the collaboration, the testing culture, and the code all point the same way.
The single most human artifact in the repo is its test methodology: a curated library of the engine's own blunders, each proven fixed.
// stupidities/cases/S-0001/case.json
"title": "Search heuristics choose non-defensive rook moves under king attack",
"status": "fixed",
"fen": "4r1k1/1p3pp1/1p1p3p/1Pr5/4Ppn1/1B1P1N1q/1RP2P2/3QR1K1 w - - 0 23",
"bad_move": "b2b1",
"reference_best_move": "b3d5",
"reference_note": "...b2b1 and b2a2 are consistently refuted by ...Rc5-h5 and
...Qxf3/Qf2 mating lines.",
"before": { "depth_moves": { "6": "b2b1", "8": "b2a2", "10": "b2a2" } },
"fix": { "summary": "Correction history now updates only after root searches,
and LMP/null move pruning are guarded in tactical king-pressure positions." }
base-before-bad.mp4 → base-after-fixed.mp4). This is painstaking, idiosyncratic human QA — the opposite of anything a language model emits on its own.Two developers, working the way open-source engines are actually built.
| # | Signal | What it shows | Weight |
|---|---|---|---|
| 1 | Two distinct authors merging via PRs (e.g. "Merge pull request #16 from starius/shrink-nnue") | Genuine multi-person open-source workflow | Decisive |
| 2 | The curated, video-documented "stupidities" regression suite | Idiosyncratic human testing methodology | Decisive |
| 3 | Quantified perf commits: "mailbox lookups in movegen and search (+23% NPS)" | Measured, incremental optimisation — a real dev loop | Strong |
| 4 | Russian personal README; casual commits ("mend", "Revert some fixes") | An individual's voice and workflow | Strong |
| 5 | Nix flake, Python training notebooks, Elo-test configs | A fully-outfitted, seriously-maintained project | Support |
| 6 | Zero AI residue or agent artifacts | No generated-code fingerprints | Support |
| Where it resembles Stockfish |
|---|
| NNUE evaluation, magic bitboards, shared transposition table |
| Modern search: SEE, NMP, LMP, correction history, quiescence |
| Syzygy tablebase support and Polyglot opening book |
| PR-based multi-developer workflow with regression tests |
| Where it differs |
|---|
| Rust vs C++; ~8k LoC and two authors vs a vast, decades-old team effort |
| An avowed learning/experimentation project rather than a title contender |
| Bespoke video-backed "stupidities" QA rather than Fishtest-scale SPRT |
Ember is a human-written, collaboratively-developed Rust engine — a personal learning project that attracted a second contributor and matured into a properly-tooled codebase. Its hand-curated, video-documented library of the engine's own blunders is the kind of painstaking, personal craftsmanship that no generator produces. It belongs firmly on the human side of the series, and stands out for the care in how it is tested.