Methodology
How the corpus was built: a three-phase pipeline (ranking → probe → validation) over the Linux drivers/ tree.
read →Phase 1 — Dormancy ranker
A no-LLM, git-log-only score that produced the shortlist of candidates worth probing.
read →Phase 2 — Codex probe
Per-driver dossier generation with strict JSON Schema output and a curated tool surface.
read →Schema reference
The dossier shape, field by field. Mirrored by the site's Zod schema; drift fails the build.
read →Calibration
Empirical results: verdict distribution, confidence histogram, cost-per-dossier, regressions.
read →Limitations
What this corpus is not. Snapshot semantics, model bias, source coverage.
read →Reproduce
How to re-run the pipeline against a new kernel SHA, swap the model, or edit the prompt — plus bulk JSON / CSV download links.
read →