Systems where AI does part of the research and a named human still answers for it. Pick a job to compare the ones that do it: what separates them is what the machine’s call is worth before you touch it.
What are you here to do?
Kept current every week
Every Monday a scheduled AI agent searches arXiv, launch announcements, the related awesome lists and GitHub for new research systems. It tests each candidate against the rules below, checks whether any excluded system has opened to the public, and opens every listed link to catch products that moved or closed. It writes a report and changes nothing. The maintainer rules on each candidate, and an entry is added or moved only when the evidence holds up.
A system has to pass two tests on its own public pages. A researcher can use it today, with no waitlist or private preview, and it was built for research work, not as a general assistant. A system that passes both and answers a question no listed system answers gets its own row. One that answers the same question as a listed system joins its cluster as one line, and each cluster’s featured system is chosen by written criteria, with arrival order counting for nothing. A university project and a company product are judged the same way.
Each system is filed by what the human still rules on, not by the stage of work it serves. The part of the product that carries research judgment is the part tested. Bench is tested first, then Delegation, Standing Verdicts, Human Approval and Critique in that order, and the first match places it. An entry moves whenever a release changes how the product works.
Every boundary is a test a reader can run against the vendor’s public pages. Pages are fetched and read directly, never through a tool that summarises them. A claim found only in page source or a script bundle carries a †, and anything no public page states is marked not documented. Each value in the How they differ tables keeps the URL and the exact words it rests on, and a check on every push rebuilds the README and this page from those sources and fails if either differs from what was committed.
The full tests, the placement rules and every exclusion are in METHOD.md. How each column is filled is in CONTRIBUTING.md, and four systems are traced through a real task in WORKFLOWS.md.
It may be here already. A system that answers the same question as a listed one appears as a close alternative, listed inside that system’s stacked card: open + close alternatives on the card it most resembles. Otherwise it usually failed one of the tests above: a waitlist, a general-purpose assistant, a language service, a plugin that only runs inside another agent, or a single maintainer with nothing behind the project. It may also have been removed on the maintainer’s judgment, or not tested yet.
Systems left out after a closer look are in the exclusion table, each with its reason. The weekly scan rules out more candidates than the table records, so a system missing from it may already have been looked at. Either way, open an issue with a link to it and the case for it.
No. It means nothing we found on the product’s public pages says either way, and features behind a login stay out of view. A ✗ appears only where the vendor says the product does not do something, or where the system’s own row in the README rules it out.
Two tools that do the same job can leave you answerable for very different things. One screening tool’s label is the record unless you open the paper and reverse it; another’s waits for a named person to rule. The jobs above let you start from the task, and the category on each card tells you what that tool leaves you to check.
Yes. Placement follows what the product documents, so a release that adds a reviewer log, a checkpoint, or a button that applies every fix at once moves the entry. The weekly scan reopens every listed product, and a pull request that re-tests an entry is as welcome as one that adds a tool.
General-purpose assistants are not built for a particular step of research, and researchers already know them. Models such as AlphaFold predict a property of nature, so they stand in for a measurement rather than for a judgment about how the research should go, which puts them outside the definition this list starts from.
Open an issue or a pull request on GitHub, and link the vendor page your claim rests on. CONTRIBUTING.md explains how each column is filled and what evidence each kind of claim needs.