ZigQueen

What it is

A from-scratch UCI chess engine in Zig, written clean-room, with a pure NNUE evaluation: a small neural network computed incrementally during search. Coding agents did much of the implementation. Architecture, experiments, evaluation and release decisions stayed with me.

The number, and why it matters

ZigQueen is independently rated at roughly 3570 Elo on the CCRL Blitz list, a third-party rating pool that has measured computer chess engines for decades.

For readers outside chess: the strongest human players in history sit below 2900. A rating near 3570 is not "strong for a hobby project". It is far beyond any human player, and the measurement is external. Nobody has to take my word for it.

The development loop

Looking right counts for nothing here. Every candidate change walks a fixed ladder:

1. Correctness gates: build, tests, stability, UCI compliance.
2. A 192-game screening match against the accepted baseline.
3. A statistical head-to-head (SPRT) against that baseline.
4. An external gauntlet against other engines as ground truth.

Training loss is only a prefilter; match play decides promotion. Version numbers are bound to the same process: a minor release is, by definition, a promoted strength gain that survived the ladder.

Calibration

The result that matters most to me is not the rating. It is that my internal strength measurements predicted the external CCRL result to within about 25 Elo. That is what a calibrated evaluation loop buys you: the right to trust your own numbers before the outside world confirms them.

The transferable lesson

It has little to do with chess. AI-generated code becomes trustworthy through measurement, not confidence. Chess happens to provide unusually clean referees; most software does not. But referees can be built: tests, benchmarks, telemetry, independent recomputation. The discipline of building them is the part that transfers to any domain, and it is the discipline this project runs on.