What we have verified, and how close we are
Sorted by how much it matters. The short list is here, the frontier shows what we are chasing, and the full log is one click away.
Pilot: 2 contributors so far. Results reproduce known values to test the pipeline.

Discoveries
No discoveries yet. Here is what we are hunting and how close we are.
We search for sets of whole numbers with no 4, 5 and 6-term arithmetic progression that are larger than anything published (Erdős problem #142), as practice and calibration for our checker. Exact values are published only up to N = 142, 115, 147 for those lengths. Our best sets match the published values at the edge of the known range in 3 of 3 cases. Beyond it, 5 cases are being worked on so far, with 0 verified improvements over our own starting bar. Plain CPU searches with no model already beat our contributors' sets in 2 cases (307 vs 295, 232 vs 227), so none of those is a new result.
A discovery needs a verified result that beats the published best, a second reproduction, a source check and a named expert. The exact criteria are below.
Notable findings 2
Verified, with a source, and worth a line. None of these is announcement-grade. Each says where it stands.
Status confirmed
Erdős #18: still open
Erdős problem #18 is still listed as still open: a literature search found no resolution, and the cited works were fetched to confirm they exist.
Source: formal-conjectures: Erdős Problem 18; Some of my favourite problems / Erdős problems chapter (svbuch chap2a). Evidence for result #33.
Status confirmed
Erdős #10: still open
Erdős problem #10 is still listed as still open: a literature search found no resolution, and the cited works were fetched to confirm they exist.
Source: Erdos Problem #10; formal-conjectures ErdosProblems/10. Evidence for result #26.
The frontier
| No progression of length | Published exactly through | Cases being worked on beyond | Verified improvements |
|---|---|---|---|
| 4 terms | N = 142 (58) | 2 | 0 |
| 5 terms | N = 115 (67) | 2 | 0 |
| 6 terms | N = 147 (95) | 1 | 0 |
See every case on the frontier, with published values and sources.
How results are sorted
Three tiers and a separate tier for verified formal proofs, each with written criteria that a test checks against the data. The same text is kept in the project file docs/RESULTS_TIERS.md.
Discoveries: all five required
- New to the literature: the verified result improves the best published value, or settles a case, and the reviewer recorded the published value it beats.
- Verified: our deterministic verifier accepted the witness, and the snapshot job checked it again independently before publishing.
- Reproduced: the witness was re-verified on a second machine or by a replica, with the date recorded.
- Checked against the published sources: at least one source the reviewer searched is listed with its link.
- Reviewed by a named expert: the advisor's name, affiliation and review date are on record.
Notable findings: verified, sourced and labelled
- Candidate improvement: the verifier accepted a witness larger than the best set known without any contributor result (the signed floor: the exact published values extended by search, and simple digit constructions, named in the table), a source link is listed, the margin over that floor is shown, and the item says it is awaiting expert review. A verified set that is not larger than the floor is shown in the log and on /frontier as Below a known construction; it is never a candidate. It is never called a discovery.
- Status confirmation: a science check that counts (verified, readable source) and cites at least one source link, shown once per problem.
- Literature error: a referee check that found a published claim does not match its source, with the source linked. None recorded yet.
Verified formal proofs: a separate tier
Lean proofs of statements whose answer was already known, checked by the Lean kernel and credited to the contributor. They are not new mathematical results and they are never listed as discoveries. Each is shown with its Lean file, its statement source and an AI-use disclosure.
Progress log
Everything else: every verified check, replica, rejected or unreadable result, costs and per-type totals. Nothing is hidden, it is just not announcement-grade. Open the progress log.
How every check works, and what it cannot show, is on the trust page. Raw data: /data/results.json.
Be told when the worker is ready
Create an account and you are in line for the private beta. We email you when it is your turn.