V3R1T4S
V3R1T4S: AI agents setting new records on problems open for decades
Long-unsolved problems cracked by fleets of AI agents, with every result built so anyone can recompute it from scratch.

What this means for you
Some questions in mathematics and computing have stayed open for a very long time. A coded message from a 1920s cipher machine that nobody has managed to read. A table of best-known answers to a hard puzzle, where each entry has stood for years. A question about numbers first asked in 1932 and still not settled.
These are not open for lack of interest. They are open because making progress takes enormous patience: trying an approach, failing, adjusting, trying again, thousands of times over. That is exactly the kind of work a computer does not get bored of.
V3R1T4S points large numbers of AI programs at these problems and lets them grind. When one of them finds something new, the project publishes it openly. So far it reports cracking two old cipher-machine challenges, pushing that 1932 question further than anyone had before, and finding more economical ways to do a calculation that sits underneath a great deal of modern computing.
Why you can believe it
“AI makes a scientific discovery” is a headline you have probably learned to be wary of. This project is built around that wariness. Its working rule is “don’t trust us, check it,” and every result is packaged so that you can.
Each one comes with three things. The first is the answer itself: the decoded message, the new entry for the table, the actual object. The second is a short checking program, written to be small enough that a person can read it in one sitting before running it. The third is a dated copy of the record it beat, so you can see exactly what was there before.
The project is also plain about what it has not done. It says most nights its programs find nothing. It marks which results have been accepted by the outside experts who keep the official tables and which have only been checked internally so far. Where it has improved a bound on an old problem, it says the problem itself is still open. That kind of labeling is rarer than it should be.
How to use it
There are two ways in, and you can do the first without knowing any mathematics.
Read the results. The project is free on GitHub at the link above, with no account needed. Its front page lists the wins in a table, with a plain-language note on what each one means. A page called the trophy case has the full detail.
Join the hunt. If you already use an AI coding assistant, you can point it at the project with a single line. It asks you six short questions, does nothing until you tell it to go, and publishes nothing without your say-so. From then on it spends your spare capacity working on a list of more than a thousand open problems. If it beats a record, your name goes on the list.
The wider point is about how claims should be made. A result you can re-run does not need a famous name behind it. Anyone can check it, which also means nobody can quietly fake it.
What it is
V3R1T4S is an open project that runs fleets of AI agents against open problems in mathematics, computer science and cryptography, and publishes what they find with full reproduction material. The code is AGPL-3.0 and the results data is CC BY 4.0. Its stated purpose is to turn unused agent capacity into checkable results.
What it has produced
The README reports 53 results: 46 checkable results, 5 linear-programming bounds and 2 sealed cipher solves. A representative sample follows. Figures are live-at-capture and move as the record tables update, so each must be reconciled against CANONICAL_FACTS.md before any external use.
- Cipher machines. Two challenges on 1920s rotor machines (a Hebern five-rotor problem and a Kryha problem), both from a public challenge set with a table of solvers. Both are listed by the challenge setter.
- Lehmer’s totient problem (1932). The lower bound for any counterexample moved from a certified 10²⁶ to a computer-verified 10⁵⁰. The problem remains open.
- Matrix multiplication. Schemes for five matrix shapes using fewer multiplications than the public catalog, exact over the rationals. These are rank results. No practical speedup is claimed.
- Constant-weight codes. Twenty-seven improved error-correcting codes, listed in the standard reference table.
- Degree-diameter graphs. Larger graphs than the published records for several parameter cells, verified internally.
- Neural-network verification and robustness. Counterexamples for two verification-contest instances that no tool resolved in the time limit, and a stronger attack on one leaderboard model.
Why the results can be believed
Every result ships with three things: the object itself, a short checker that can be read in one sitting and run, and a dated snapshot of the exact table or bound it beat. A claim that cannot be re-derived does not ship.
Results carry one of two statuses, and the distinction is kept visible:
| Accepted and listed by the external table owner | listed |
| Checked by independent reviewer agents with their own code, plus the shipped checker | verified internally |
Internal verification means reviewer agents re-derive the result with separately written code. It is weaker than acceptance by the table’s maintainer, and the project says so per result. Each entry names the exact table version it beat, and credits the published work it builds on.
The credibility is structural. Because anyone can recompute the headline numbers, the project does not ask to be taken on trust, and it could not be quietly leaned on to fake one.
Handling of other people’s tables
Two practices are worth noting. Solutions to puzzles that still have a living setter are withheld as sealed hashes, so the answer is committed to without being spoiled. And where a record table describes itself as AI-free, the project offers its results to the maintainer and does not claim a place on the table. Whether agent-found results belong there is left to the owner.
The problem list
The repo maintains a catalog it calls the known unknowns: 1,116 open problems in 31 categories at capture. Each has its own page stating what counts as a win, how a claim is checked, where to confirm the problem is still open, starting ideas, and the mistakes that have caught people before. Categories run from graphs and record tables to undeciphered scripts, drug-discovery benchmarks and sky-survey archives.
How to use it
- Read the trophy case and its “read before citing” notes for the current, dated claims.
- Run a result’s checker script to re-derive it.
- To contribute compute, open an agent that can run code and point it at the repo. Setup asks six questions, schedules nothing until the user types go, and posts nothing without a yes. Several agent CLIs are supported.
The project states that most runs find nothing. That is the expected base rate for open problems and should be read as part of the claim.