Release notes by merge date. For what is planned, see the roadmap on the MCP docs.
2026-10-06
Pre-submission check, author loop, categories
Ask AutoXiv at /ask (and ask_corpus over MCP): a question answered across the corpus from the papers’ overviews and evidence, with citations that open the paper and a plain “nothing relevant” when the corpus has nothing.
Reproducibility Check at /check: a PDF or arXiv id plus a repository → extracted claims, a repository inspection, numbers cross-checked against the results files, and a 0–100 score with concrete fixes — before anything is public.
Verified authors: claim a byline with a signal the server can check (ownership of the registered repository, or an email match); claimed bylines show a Verified author chip and feed the dashboard.
Author notifications: one transactional email when a reproduction reaches a verdict on your paper, when an arXiv registration goes live, or when a bounty claim is decided — opt-out on the profile page, never sent twice.
Scheduled re-verification: verified papers are re-run on a daily cron with backoff, so "last verified" stays current (ships off until the sandbox is enabled).
Alpha Machine executes its backtest scaffold in a sandbox on synthetic data, repairs it once if it crashes, and reports ran / repaired / failed / skipped with the captured output — a validated scaffold, not a validated strategy (needs the sandbox; otherwise skipped).
Category index and listings with human names (cs.LG → Machine Learning), evidence chips, per-category RSS feeds, share cards and sitemap entries; category chips on paper pages link there.
Faster first paint on the home and explore pages: corpus-wide counts and evaluation stats are cached for a few minutes instead of recomputed per request.
Transparency page at /stats (and /api/stats): papers, code share, executed runs, verdict breakdown, claim-ledger tallies, per-category table and a twelve-month series — with the exclusions spelled out.
Bulk arXiv import at /submit/arxiv/bulk: search arXiv by author or paste ids, see what is already on AutoXiv, register the rest in one go inside the daily limit.
About page rebuilt on the design system; share cards for author profiles; "With code" views are exact; a security and a correctness review of every new surface, with their findings fixed.
2026-10-06
Redesign, verification surfaces, agent tools
New type system (Source Serif 4 · IBM Plex Sans · IBM Plex Mono), navigation, and first-glance surfaces; secondary pages on the same system.
Verification ladder under every paper header; README badge at /abs/<id>/badge.svg; per-paper share cards; ScholarlyArticle metadata; RSS feed.
Ask this paper: evidence-grounded Q&A with question-aware retrieval over the full text (page-anchored when a parse exists).
Register any arXiv paper by link — metadata imported, PDF linked (not copied), optional repository, overview pipeline — on the web and over MCP (register_arxiv_paper, ask_paper).
Explore pagination and category sidebar; For-labs page; Terms and Privacy; ORCID links on author chips.
Security hardening: reproductions run only on the registered repository; visibility checks on chats, workspaces, and runs; metered LLM endpoints; BYOK fails closed; sanitised search filters.
2026-09-01
Verification release
Typed claim ledger (claim.v1): claims extracted per block, re-extracted blind, agreement decided by code; only confirmed claims can be scored — enforced in the database.
Server-captured run manifests, structured result channel, executor/judge separation with a leakage check.
Multi-seed replicates with Wilson / BCa intervals for stochastic claims; budget tiers; two-phase sandbox egress control.
Demo evidence purged; verdicts clamped to observed evidence; match scores server-computed; bounty settlement paused pending escrow.
2026-08-26
MCP v0.2–v0.5, Open Calls
Alpha Machine and reproductions over MCP; list_bounties, get_paper_claims; usage dashboard on the API-keys page.
Open Calls: every OpenReview venue accepting submissions, nearest deadline first, with a prefilled openreview-py snippet.
Evaluation scorecards on paper pages; live MCP docs with worked sessions and an in-browser call.
2026-08-20
Marketplace and Alpha Machine
Reproduction bounty marketplace: pooled stakes per paper, claims linked to reproduction runs.
Alpha Machine: paper → alpha screen → code discovery → strategy spec → backtest scaffold, with downloadable deliverables and a grounded follow-up chat.
Workspaces with roles and invitations, agent chat with paper context, lab assistant.
2026-04-26
Reproduction agent and MCP v0.1
Sandboxed reproduction agent: clone, install with recovery, run, parse metrics, compare to claims, structured verdict; public replay of completed runs.
First-party MCP server with read tools, library/review/cluster-chat writes, trial tier and per-key limits.