Research papers
Reproducible safety benchmarks of raw LLM fitness coaching versus the WPL governance layer.
How these papers work
This is a versioned research program, not a single paper. Each paper version is pinned to a tagged release of the open wpl-eval corpus: the headline numbers, model lineup, and scenario count in a paper reflect exactly one corpus tag, and every quantitative claim is offline-reproducible from the committed raw model outputs — no further API spend required.
Revisions follow the corpus release cadence rather than the calendar. We publish the full lineage here — including superseded drafts — so the progression of the work is visible: what each version measured, what it got wrong, and what the next version fixed. The current version is prepared for arXiv submission; until it is up on arXiv, this page is the canonical home of the papers.
Version history
Reproduce the results
Every prompt, model output, score, and the deterministic scorer are published at github.com/gymbile/wpl-eval, with each paper pinned to its release tag. An interactive view of the benchmark lives on the eval page.