Source Authorship Forensics
A source-level read of Tors3/Triumviratus v4.2 — the most nuanced case in the series.
Triumviratus is the work of a serious engine developer applying rigorous empirical method over two and a half years: SPRT-verified search changes, multi-thousand-iteration SPSA parameter co-tunes, A/B NPS tests, anchored Elo gauntlets — "nothing is merged on feel." That is a human development loop of a high order, documented commit by commit.
The repository does contain a CLAUDE.md — the developer uses the Claude Code assistant (with a code-graph navigation tool) as part of their workflow. But here AI is a tool inside an expert's years-long process, not the origin of the code. Two honest caveats sit beside the verdict: the work is AI-assisted, and the engine is a Stockfish-derived lineage (see below).
Confidence: High on human authorship; the AI-assist and Stockfish-derivation are noted for full transparency.
The release notes read like an experimental logbook. Each gain is measured, with confidence intervals.
| Evidence | What it shows |
|---|---|
| 116 commits over ~2.5 years (Jan 2024 → Jun 2026), one named author | Sustained, long-horizon personal development |
| "Large SPSA co-tune (~10k iterations, 18 parameters)"; "mega-SPSA (55 parameters)" | Serious automated parameter tuning on real hardware/cloud |
| "SPRT-verified vs 4.0 … +35.0 ± 17.9 Elo at 20+0.2 (LOS ~100%)" | Statistically-disciplined testing, not vibes |
| clang-PGO release builds; interleaved A/B NPS tests; bench fingerprints | Performance engineering at a competitive level |
| Reverts, version-banner fixes, dead-code cleanups | An organic, human, sometimes-messy workflow |
Two engines in this series have Claude in the repository. They are opposite cases.
// source/CLAUDE.md — instructions for the Claude Code assistant "This project has a knowledge graph at graphify-out/ ... For codebase questions, first run `graphify query ""` ... After modifying code, run `graphify update .` to keep the graph current."
CLAUDE.md configures an AI assistant to navigate a large existing codebase (via a code-graph tool) — it presupposes a substantial, already-built engine to reason about. Contrast Owen, where a committed claude dependency sat beside an entire engine delivered in a single 11,000-line commit. Same assistant; here it augments an expert's long-running project, there it produced the project. AI assistance of this kind is now normal and legitimate; it does not make the work machine-authored.More than resembling Stockfish, Triumviratus is openly built on parts of it.
| Shared / derived from Stockfish |
|---|
Ships Stockfish's NNUE code (source/sfnnue/) and used SF's net nn-b1a57edbea57 as default until v4.2 |
| "SF-style" LMP and depth-pruning blocks; ProbCut; singular extensions |
| Lazy SMP with a shared transposition table |
| SPRT/SPSA testing culture (the Fishtest methodology) |
| Where it is its own work |
|---|
| Original single-board NNUE bridge feeding the net from native bitboards |
| Bespoke 24-byte TT with 16-bit static eval; staged MovePicker; correction history |
| Its own large SPSA-tuned search vector (the real source of its Elo) |
v4.2 "first own-lineage": begins shipping its own rubicon network |
The Stockfish derivation is a question of originality, not authorship — and the developer discloses it openly (SF-net default, GPLv3 source-link notes). It remains substantial, expert human work layered on a GPL lineage.
Triumviratus is human-written — and among the most methodologically serious engines in this series — developed by one expert over two and a half years of SPRT- and SPSA-disciplined work. It is also the series' clearest example of legitimate AI-assisted development: a CLAUDE.md shows an AI assistant in the loop, but as a tool augmenting a person's long-running project, the polar opposite of an agent-generated dump. Two fair caveats travel with the verdict — the work is AI-assisted, and the engine is a Stockfish-derived lineage only now growing its own net — but neither moves it off the human side of the line.