Direct building · tests · failure boundaries
I direct substantially more code than I hand-write. I own the architecture, decomposition, tests, failure boundaries, and the decision that a result is ready to trust. The artifacts are the useful evidence.
Claims, checks, caveats, reproduction commands, and the places where the mechanisms lose. It is a verification record rather than a résumé argument.
Read the record →A heavy build arm fits semantic, structural, and cognitive-move dimensions against a corpus; a dependency-free read arm projects live prompts and performs advisory routing. The six-axis subset retained ~88.4% of candidate-pool variance in the bundled corpus, while heavy vs pure routing held 20/20 parity.
GitHub →Maps up to six scalar dimensions into luminance, red/green, and blue/yellow opponent channels with null-aware bilinear interpolation. Palette luminance is constrained so a pattern that vanishes under palette changes is treated as a rendering artifact, not promoted structure.
Read the method →A small line-oriented transform language emits paired forward/inverse Rust kernels at build time. CI checks generated-source drift and round-trip behavior; top-level STORE/zlib/LZMA baselines retain veto power over custom transforms.
GitHub →A from-scratch classical chess engine with perft-correct move generation, toggleable ordering/mechanism cells, ablation and match harnesses, structured per-decision telemetry, and a separate comparison/viewer path against Stockfish.
GitHub →The workflow is see → name → permute → discriminate → derive → subtract → look again. A visible candidate must become an exact predicate or measurable structure before promotion. Unsupported cells remain null rather than being smeared into numeric zero.
