anzenresearch/cybench-oc3-reports

Rebuild all seven per-model run reports from a single template: 7 publication-grade figures each, consistent structure, human-vs-model difficulty analysis, placement stability, clustered uncertainty, transcript availability. Removes base64-bloated HTML and legacy hand-rolled SVGs.

Restructure as a standalone measurement audit: add summary, contributions and reproducibility sections, separate pipeline findings from study limitations, institutional typography, fix perfect-score count.

Standalone OC3 audit: recompute risk model live, quantify KRI threshold fragility and FST-ordering violations, propagate placement uncertainty, add clustered intervals and judge-reliability floor, eight new figures.

Upload cybench-oc3-spar-reports artifacts

initial commit