ARR exists because uploading a PDF proves almost nothing. It distinguishes work that has survived disclosed frontier-model attacks on an exact version from work that has merely been posted online.
A versioned registry where new research must survive a disclosed hostile frontier-model audit selected for that assessment round and a human decision. Canonical manuscripts, prompts, model identities, findings, code, provenance and verification records remain inspectable.
Research papers present complete scholarly arguments at paper scale. Technical notes preserve narrower but rigorous results, proofs, formalizations, methods, replications, negative results, software or protocols. A note is different in scope, not exempt from evidence or integrity requirements.
ARR is not a journal, a replacement for expert peer review or a guarantee that a scientific claim is true. Activity rankings measure use; the separate scientific ranking reports version-locked model opinions with their provenance and limits.
Lluis Eriksson is founder, registry operator, responsible editor and data controller. Every decision is human. His conflicted or author-owned work requires a disclosed independent editor before publication.
Stable identifiers are independent of GitHub. Versioned releases distribute generated and large files; future object storage and independent preservation mirrors can replace any provider without changing citations.
ARR does not currently charge for submission, assessment, publication or withdrawal. Authors use one direct private form without requesting an invitation. Manuscripts enter a separate quarantine service; GitHub and ordinary email are never manuscript channels.
ARR deliberately uses the strongest available frontier models as tireless adversarial referees. Published assessments remain comparable over time: model, date, score, strengths, weaknesses, possible errors and novelty candidates are preserved rather than overwritten.