Machine-readable findings: correlations + per-model summary + caveats (#4)

Prose analysis: per-model table + arch-feature correlations (#3)

Dataset rows: 4 models, arch features + zero-shot scores (#2)

Add dataset card: method, panel, honest finding, caveats (#1)

initial commit