New reproducibility push: Claw + SKILL.md conference
Stanford and Princeton launched a conference requiring AI papers to be fully reproducible by an AI called Claw via a single SKILL.md file — the contest carries $50K prizes and aims to tackle the reproducibility crisis. (x.com)
Claw4S 2026 requires authors to submit an executable SKILL.md that the Claw agent can run end-to-end instead of (or alongside) a static PDF, making "executability" the primary acceptance criterion. (theaivalley.com) Organizers list Stanford and Princeton as hosts and have named conference chairs including Le Cong (Stanford) and Mengdi Wang (Princeton), with Yingcheng (Charles) Wu and Zhe Zhao on the organizing team. (t.co) The event is running a prize pool of roughly $50,200 with a published cap of up to 364 winners and an April 5, 2026 submission deadline for hackathon-style entries. (github.com) Submissions are ingested via a clawRxiv API (POST /api/posts) that accepts title, abstract, Markdown content and an attached skill_md file intended to be machine-executable by Claw. (github.com) clawRxiv’s tooling includes an automated ClawReviewer that runs a multi-stage pre-screen with deterministic static checks (14 checks listed) covering SKILL.md structure, dependency analysis, and step-chain integrity before any dynamic execution. (clawrxiv.io) An independent archive audit of the clawRxiv snapshot found only 1 cold-start-executable skill_md among 34 archived artifacts, prompting work like "SkillCapsule" to automatically repair common failures and improve first-contact reproducibility. (clawrxiv.io)
