Four problem pages, the lab, the atlas, and the renders, tied together: everything the program found, and everything it did not.
Seven rounds, November 20-21, 2026. 66 receipts, hash-chained, head
bdebd1e7d93dd108bffe0ba01d499e603db81a8c17027bb0aad36a70f73dacdd.
Nothing is solved. Each problem below has its own page with everything in it:
the statement, the method, the round-by-round numbers, the figures, the
receipts, the limits, and the next attack.
First 4,000 zeros re-verified on the critical line, two independent implementations, max real-part deviation 0.0 at 30 digits. Re-verification, not a new record. Read the page (rounds 1, 2, 4, 5).
130 elliptic curves, conductor to 50: 128 rank 0 with L(E, 1) nonzero, 2 rank 1 (37a1, 43a1) with L(E, 1) = 0 and L'(E, 1) nonzero. The pattern holds on every curve checked. Read the page (rounds 1, 2, 4, 5).
One concrete decision problem from a real codebase. Registry search in P (0.0718 s). The defeater-free fragment in P by greedy search (44 of 44 against brute force). The full fragment with defeaters also in P (8,876 checks, 0 violations). The hardness direction stalls on structure, and existence is 1,400 of 1,400 at every density. Read the page (rounds 1-5).
Formalized in Lean 4.13.0: 58 theorems, 0 sorries. The Lefschetz (1,1) case is now conditional on exactly two named inputs (Dolbeault isomorphism, Hodge decomposition), with the sheaf long exact sequence proved for the actual exponential sequence. Read the page (rounds 1-5).
These two have no standalone page yet; their rounds live on the lab page itself, with the field renders alongside. Navier-Stokes: full 963 at nu = 0.000125, Kida-Pelz growth 145.79, Taylor-Green 138.27, both smooth, no blowup, no saturation; Taylor-Green to T = 20 is transient too. Yang-Mills: the torelon channel is overturned (three seeds disagree, no stable plateau), so no continuum extrapolation is attempted. Read the lab (rounds 1-7).
| Item | Where |
|---|---|
| The lab, all seven rounds | index.html |
| Claim landscape, all six problems | atlas.html |
| Riemann page | riemann.html |
| Birch and Swinnerton-Dyer page | bsd.html |
| P versus NP page | p-vs-np.html |
| Hodge page | hodge.html |
| Field renders, rounds 1-7 | renders/ (r1 to r7, 1080x1350, computed from the lab's own solvers) |
| Round data and receipts | data/ (all JSONs, receipts.json, lab.db) |
| Lean sources | lean/ (Hodge/*.lean, axiom audit logs) |
| Solver and receipt scripts | compute/ |
Three overturns and one reduction to named gaps. The Yang-Mills torelon channel was overturned twice: first by a second seed, then by the estimator test, which showed the mass moves more across tau than across chains. The Hodge round 2 assessment that O_X was not tractable was overturned by round 3, which built it. The Hodge (1,1) case was reduced to exactly two named missing inputs. The P versus NP phase transition was shown not to survive defeaters. And the calibration work held without a break: 4,000 Riemann zeros on the line, 130 BSD curves following the rank prediction.
No solution to any of the six problems. No proof, no disproof, no counterexample. No new record on Riemann (re-verification only). No BSD formula check. No hardness reduction for P versus NP. No Hodge proof beyond the (1,1) scaffolding. No Yang-Mills continuum claim, because the channel that would carry it failed. That is the honest state of the program after seven rounds.