Neuruh · Millennium Problems Program

The Millennium Lab

Round seven of the program. Real computations, run here on November 21, 2026, receipted in the sovereign runtime hash chain. No solution claims.

Date
November 21, 2026
Runs
3 of 3 moves run, round 7 complete
Receipts
66 in chain, valid after restart
Chain head
bdebd1e7d93dd108...
Honesty block. Nothing on this page claims a solution to any Millennium Problem. The Riemann run re-verifies known ground: the first 4,000 zeros were already known to lie on the critical line, and this run checks that fact again with an independent implementation. The BSD run computes ranks and L-values for one small curve family; 128 of 130 curves are rank 0 and two are rank 1, which is also known ground. The hunts are attacks on open questions, not solutions: a blowup search that stayed smooth at still lower viscosity, a polynomial-time search for a wider fragment of one complexity object, a five-point lattice extrapolation that got weaker when the fifth point arrived, a variational glueball measurement at a fifth spacing, and a Lean formalization of the sheaf long exact sequence with c1, with the Dolbeault isomorphism still blocked. Round 6 ran two moves in sequence: a Navier-Stokes deep dive at nu 0.000125, where enstrophy growth rose again to 88.24 and stayed smooth, and a Yang-Mills second seed that overturned the round 5 torelon points, so the five-point extrapolation cannot support any continuum claim. Round 7 ran three moves in sequence: the full 96^3 Navier-Stokes run at nu 0.000125, where enstrophy growth rose again to 145.79 with no saturation, the longer-time Taylor-Green to t = 20, which peaked and decayed like Kida-Pelz, and a Yang-Mills third seed that showed the torelon channel has no stable plateau, so the overturned verdict stands. Field renders for rounds 5, 6 and 7 were computed from the lab's own solvers. A computation is reported only with the numbers its own run produced, on the date it ran. Every number on this page comes from a run on November 20 or 21, 2026. Note: on November 20, 2026 the founder ordered the hosting repository NeuruhAI/neuruh-demos flipped to private. This page is now hosted publicly at https://millennium-problems-6a8.pages.dev/lab.html.

01What the lab is

The atlas stated the six problems and named six computable first moves. The lab is where those moves get executed. Each move is a real program with a real run, a real result file, and a hash-chained receipt, so anyone can re-run it and check it. Round 1 ran the two calibration moves and named the four hunts. Round 2 ran all four hunts, plus calibration extensions: BSD to conductor 30 and Riemann to 2,000 zeros. Its headline results: Yang-Mills three-point torelon extrapolation intercept 3.473 plus or minus 1.506 with the slope error 500 times the slope, supporting nothing; Navier-Stokes Kida-Pelz at 643 grew enstrophy 18.86 and stayed smooth, growth saturating with resolution; P versus NP defeater-free fragment decided in polynomial time by a node-wise greedy search; Hodge group-level exponential sequence formalized, sheaf O_X assessed not tractable. Rounds 3 and 4 ran simultaneously, nine moves in total: four hunts in each round plus a round 4 calibration extension. Round 5 ran five moves: the fifth beta point at 2.6, lower viscosity at nu 0.00025, the wider P versus NP fragment, the sheaf long exact sequence with c1, and calibration extensions to conductor 50 and 4,000 zeros. Round 6 ran two moves in sequence: the Navier-Stokes deep dive to nu 0.000125 and the Yang-Mills second seed at beta 2.5 and 2.6. Round 7 ran three moves in sequence: the full 96^3 Navier-Stokes run at nu 0.000125, the longer-time Taylor-Green to t = 20, and the Yang-Mills third seed at beta 2.5 and 2.6, plus field renders for rounds 5, 6 and 7. A move counts only when its receipts exist.

The receipt chain is the sovereign runtime receipt format, direct reuse: the production SovereignStore class from ~/workspace/sovereign-runtime, canonical JSON, SHA-256, each receipt carrying the hash of the previous one, GENESIS head. The lab ledger is a separate store (data/lab.db) in the same format. The chain was verified after the store was closed and reopened, which is the restart proof. Full chain head: bdebd1e7d93dd108bffe0ba01d499e603db81a8c17027bb0aad36a70f73dacdd.

02Riemann: first 4,000 zeros, re-verified

RUN November 21, 2026, round 5 calibration. Method: PARI/GP 2.19.1 lfunzeros locates every zero with imaginary part in [0, 4650] (C implementation, Riemann-Siegel, argument principle). Then, for each located t, an independent mpmath Newton step from 0.5 + i t using zeta and zeta prime, at 30 decimal digits. The Newton step is free to move off the line; the reported real part is where it converges. The two implementations agree: the first zero from PARI matches the mpmath bisection value to 22 digits.

Result: 4000 zeros checked, first at t = 14.1347, last at t = 4506.3115. Largest deviation of a zero real part from 1/2: 0.0 at 30-digit precision (no zero deviated: every Newton step returned Re = 1/2 exactly at this precision, so the deviation is below 1e-30). Largest |zeta(1/2 + i t)| at a located zero: 7.4547e-21. Count check against the Riemann-von Mangoldt formula at T: expected 3999.53 zeros, found 4,000, so S(T) = 0.469, inside the small range the formula allows. The first 3,000 t values match round 4 exactly. The new interval [3545, 4650] holds 984 zeros, and with the 16 leftover zeros from the round 4 upper scan that is 1,000 new zeros beyond round 4. The new scan ran twice with identical hashes; the second run took 414.0 seconds and the mpmath verification pass took 560.91 seconds. The full [0, 4650] scan overflows the default 8 MB PARI stack, so the interval was split, the same workaround recorded in round 2.

This is a re-verification, not a new record. The first 1013 zeros were already verified on the line by others. This run checks the first 4,000 again, independently, on this machine, with receipts.

nt (imaginary part)Re(s) from NewtonRe deviation from 1/2|zeta(1/2 + i t)|
114.134725141734693790457250.50.01.5733e-24
221.022039638771554992628480.50.04.6167e-25
325.010857580145688763213790.50.01.3615e-24
430.42487612585951321031190.50.03.22e-24
532.935061587739189690662370.50.01.4318e-24
637.586178158825671257217760.50.06.7404e-24
740.918719012147495187398130.50.04.5991e-24
843.327073280914999519496120.50.03.9703e-24
948.005150881167159727942470.50.04.3112e-24
1049.773832477672302181916780.50.06.6386e-24

All 4,000 zeros are listed in the run file, not only the first 10 shown here. Results file SHA-256: 720c758471012ad6107678bffe23bd80b9d73bbb9ef57d7cacd9a85a89a3c73e. BSD results file SHA-256: 220e42edc74388a62fa3b8fa4a80708f6bc95c641fe0edf5fada0917e13dc92c.

03Birch and Swinnerton-Dyer: one curve family, computed

RUN November 21, 2026, round 5 calibration. Family: every elliptic curve over Q of conductor at most 50, as enumerated by PARI ellsearch. That is 130 curves, 93 from round 4 plus 37 new at conductors 41 to 50. Method: PARI/GP 2.19.1, ellanalyticrank at 30 digits of precision. The conductor of every curve was recomputed here with ellglobalred and matches its label.

Result: 128 curves have analytic rank 0 with L(E, 1) nonzero, and two curves have analytic rank 1. The rank 1 curves are 37a1 and 43a1. The 43a1 curve is new this round, model y^2 + y = x^3 + 1x^2, and its recorded L value 0.3435239746184782 is L'(E, 1), the derivative, not L(E, 1), because L(E, 1) vanishes at rank 1, which is exactly what Birch and Swinnerton-Dyer predicts. The 37a1 L value matches round 4 exactly. L-values repeat within an isogeny class, because isogenous curves share an L-function. The smallest |L(E, 1)| among the rank 0 curves is 0.2538418608559107, the curve 11a1, unchanged from rounds 1, 2, and 4. No curve breaks the round 1 pattern. Run time: 0.41 seconds. The scan was run twice, and both runs produced identical hashes.

LabelMinimal modelCond.RankL(E, 1)Torsion
11a1y^2 + y = x^3 - 1x^2 - 10x - 201100.25384186085591075
11a2y^2 + y = x^3 - 1x^2 - 7820x - 2635801100.25384186085591071
11a3y^2 + y = x^3 - 1x^21100.25384186085591075
14a1y^2 + xy + y = x^3 + 4x - 61400.330223659344480566
14a2y^2 + xy + y = x^3 - 36x - 701400.330223659344480566
14a3y^2 + xy + y = x^3 - 171x - 8741400.330223659344480562
14a4y^2 + xy + y = x^3 - 1x1400.330223659344480566
14a5y^2 + xy + y = x^3 - 2731x - 551461400.330223659344480562
14a6y^2 + xy + y = x^3 - 11x + 121400.330223659344480566
15a1y^2 + xy + y = x^3 + 1x^2 - 10x - 101500.350150760583150538
15a2y^2 + xy + y = x^3 + 1x^2 - 135x - 6601500.350150760583150534
15a3y^2 + xy + y = x^3 + 1x^2 - 5x + 21500.350150760583150538
15a4y^2 + xy + y = x^3 + 1x^2 + 35x - 281500.350150760583150538
15a5y^2 + xy + y = x^3 + 1x^2 - 2160x - 395401500.350150760583150532
15a6y^2 + xy + y = x^3 + 1x^2 - 110x - 8801500.350150760583150532
15a7y^2 + xy + y = x^3 + 1x^2 - 80x + 2421500.350150760583150534
15a8y^2 + xy + y = x^3 + 1x^21500.350150760583150534
17a1y^2 + xy + y = x^3 - 1x^2 - 1x - 141700.386769938387780074
17a2y^2 + xy + y = x^3 - 1x^2 - 6x - 41700.386769938387780074
17a3y^2 + xy + y = x^3 - 1x^2 - 91x - 3101700.386769938387780072
17a4y^2 + xy + y = x^3 - 1x^2 - 1x1700.386769938387780074
19a1y^2 + y = x^3 + 1x^2 - 9x - 151900.45325324449610363
19a2y^2 + y = x^3 + 1x^2 - 769x - 84701900.45325324449610361
19a3y^2 + y = x^3 + 1x^2 + 1x1900.45325324449610363
20a1y^2 = x^3 + 1x^2 + 4x + 42000.470729190326518946
20a2y^2 = x^3 + 1x^2 - 1x2000.470729190326518946
20a3y^2 = x^3 + 1x^2 - 36x - 1402000.470729190326518942
20a4y^2 = x^3 + 1x^2 - 41x - 1162000.470729190326518942
21a1y^2 + xy = x^3 - 4x - 12100.451115405388492048
21a2y^2 + xy = x^3 - 49x - 1362100.451115405388492044
21a3y^2 + xy = x^3 - 39x + 902100.451115405388492048
21a4y^2 + xy = x^3 + 1x2100.451115405388492044
21a5y^2 + xy = x^3 - 784x - 85152100.451115405388492042
21a6y^2 + xy = x^3 - 34x - 2172100.451115405388492042
24a1y^2 = x^3 - 1x^2 - 4x + 42400.53912891187491098
24a2y^2 = x^3 - 1x^2 - 24x - 362400.53912891187491094
24a3y^2 = x^3 - 1x^2 - 64x + 2202400.53912891187491094
24a4y^2 = x^3 - 1x^2 + 1x2400.53912891187491094
24a5y^2 = x^3 - 1x^2 - 384x - 27722400.53912891187491092
24a6y^2 = x^3 - 1x^2 + 16x - 1802400.53912891187491092
26a1y^2 + xy + y = x^3 - 5x - 82600.51557665127729453
26a2y^2 + xy + y = x^3 - 460x - 38302600.51557665127729451
26a3y^2 + xy + y = x^32600.51557665127729453
26b1y^2 + xy + y = x^3 - 1x^2 - 3x + 32600.62096534954905557
26b2y^2 + xy + y = x^3 - 1x^2 - 213x - 12572600.62096534954905551
27a1y^2 + y = x^3 - 72700.58887958342848333
27a2y^2 + y = x^3 - 270x - 17082700.58887958342848331
27a3y^2 + y = x^32700.58887958342848333
27a4y^2 + y = x^3 - 30x + 632700.58887958342848333
30a1y^2 + xy + y = x^3 + 1x + 23000.5586580432069166
30a2y^2 + xy + y = x^3 - 19x + 263000.55865804320691612
30a3y^2 + xy + y = x^3 - 14x - 643000.5586580432069162
30a4y^2 + xy + y = x^3 - 69x - 1943000.5586580432069166
30a5y^2 + xy + y = x^3 - 289x + 18623000.5586580432069166
30a6y^2 + xy + y = x^3 - 334x - 23683000.5586580432069164
30a7y^2 + xy + y = x^3 - 5334x - 1503683000.5586580432069162
30a8y^2 + xy + y = x^3 - 454x - 5443000.5586580432069162
32a1y^2 = x^3 + 4x3200.65551438857302994
32a2y^2 = x^3 - 1x3200.65551438857302994
32a3y^2 = x^3 - 11x - 143200.65551438857302992
32a4y^2 = x^3 - 11x + 143200.65551438857302994
33a1y^2 + xy = x^3 + 1x^2 - 11x3300.74733914774274364
33a2y^2 + xy = x^3 + 1x^2 - 6x - 93300.74733914774274362
33a3y^2 + xy = x^3 + 1x^2 - 146x + 6213300.74733914774274364
33a4y^2 + xy = x^3 + 1x^2 + 44x + 553300.74733914774274362
34a1y^2 + xy = x^3 - 3x + 13400.74927722105228396
34a2y^2 + xy = x^3 - 43x + 1053400.74927722105228396
34a3y^2 + xy = x^3 - 103x - 4113400.74927722105228392
34a4y^2 + xy = x^3 - 113x - 3293400.74927722105228392
35a1y^2 + y = x^3 + 1x^2 + 9x + 13500.70291123913490553
35a2y^2 + y = x^3 + 1x^2 - 131x - 6503500.70291123913490551
35a3y^2 + y = x^3 + 1x^2 - 1x3500.70291123913490553
36a1y^2 = x^3 + 13600.70109105266272716
36a2y^2 = x^3 - 15x + 223600.70109105266272716
36a3y^2 = x^3 - 273600.70109105266272712
36a4y^2 = x^3 - 135x - 5943600.70109105266272712
37a1y^2 + y = x^3 - 1x3710.30599977383405231
37b1y^2 + y = x^3 + 1x^2 - 23x - 503700.72568106193615283
37b2y^2 + y = x^3 + 1x^2 - 1873x - 318333700.72568106193615281
37b3y^2 + y = x^3 + 1x^2 - 3x + 13700.72568106193615283
38a1y^2 + xy + y = x^3 + 9x + 903800.63021074331409953
38a2y^2 + xy + y = x^3 - 86x - 24563800.63021074331409951
38a3y^2 + xy + y = x^3 - 16x + 223800.63021074331409953
38b1y^2 + xy + y = x^3 + 1x^2 + 13800.81924563305510375
38b2y^2 + xy + y = x^3 + 1x^2 - 70x - 2793800.81924563305510371
39a1y^2 + xy = x^3 + 1x^2 - 4x - 53900.82668785067338564
39a2y^2 + xy = x^3 + 1x^2 - 69x - 2523900.82668785067338562
39a3y^2 + xy = x^3 + 1x^2 - 19x + 223900.82668785067338564
39a4y^2 + xy = x^3 + 1x^2 + 1x3900.82668785067338562
40a1y^2 = x^3 - 7x - 64000.74220623671119324
40a2y^2 = x^3 - 107x - 4264000.74220623671119322
40a3y^2 = x^3 - 2x + 14000.74220623671119324
40a4y^2 = x^3 + 13x - 344000.74220623671119324
42a1y^2 + xy + y = x^3 + 1x^2 - 4x + 54200.86886186437420498
42a2y^2 + xy + y = x^3 + 1x^2 - 84x + 2614200.86886186437420498
42a3y^2 + xy + y = x^3 + 1x^2 - 104x + 1014200.86886186437420494
42a4y^2 + xy + y = x^3 + 1x^2 - 1344x + 184054200.86886186437420494
42a5y^2 + xy + y = x^3 + 1x^2 - 914x - 109154200.86886186437420492
42a6y^2 + xy + y = x^3 + 1x^2 + 386x + 12774200.86886186437420492
43a1y^2 + y = x^3 + 1x^24310.34352397461847821
44a1y^2 = x^3 + 1x^2 + 3x - 14400.80464628757465183
44a2y^2 = x^3 + 1x^2 - 77x - 2894400.80464628757465181
45a1y^2 + xy = x^3 - 1x^2 - 54500.92159087663963162
45a2y^2 + xy = x^3 - 1x^2 - 45x - 1044500.92159087663963164
45a3y^2 + xy = x^3 - 1x^2 - 720x - 72594500.92159087663963162
45a4y^2 + xy = x^3 - 1x^2 - 90x + 1754500.92159087663963164
45a5y^2 + xy = x^3 - 1x^2 - 1215x + 166004500.92159087663963164
45a6y^2 + xy = x^3 - 1x^2 + 315x + 10664500.92159087663963162
45a7y^2 + xy = x^3 - 1x^2 - 19440x + 10481354500.92159087663963162
45a8y^2 + xy = x^3 - 1x^2 - 990x + 227654500.92159087663963162
46a1y^2 + xy = x^3 - 1x^2 - 10x - 124600.66090411132397192
46a2y^2 + xy = x^3 - 1x^2 - 170x - 8124600.66090411132397192
48a1y^2 = x^3 + 1x^2 - 4x - 44800.8428751774062984
48a2y^2 = x^3 + 1x^2 - 64x - 2204800.8428751774062982
48a3y^2 = x^3 + 1x^2 - 24x + 364800.8428751774062988
48a4y^2 = x^3 + 1x^2 + 1x4800.8428751774062982
48a5y^2 = x^3 + 1x^2 - 384x + 27724800.8428751774062984
48a6y^2 = x^3 + 1x^2 + 16x + 1804800.8428751774062988
49a1y^2 + xy = x^3 - 1x^2 - 2x - 14900.96665585280840582
49a2y^2 + xy = x^3 - 1x^2 - 37x - 784900.96665585280840582
49a3y^2 + xy = x^3 - 1x^2 - 107x + 5524900.96665585280840582
49a4y^2 + xy = x^3 - 1x^2 - 1822x + 303934900.96665585280840582
50a1y^2 + xy + y = x^3 - 1x - 25000.71316498142829683
50a2y^2 + xy + y = x^3 - 126x - 5525000.71316498142829681
50a3y^2 + xy + y = x^3 - 76x + 2985000.71316498142829683
50a4y^2 + xy + y = x^3 + 549x - 22025000.71316498142829681
50b1y^2 + xy + y = x^3 + 1x^2 - 3x + 15000.9568112265876285
50b2y^2 + xy + y = x^3 + 1x^2 + 22x - 95000.9568112265876285
50b3y^2 + xy + y = x^3 + 1x^2 - 13x - 2195000.9568112265876281
50b4y^2 + xy + y = x^3 + 1x^2 - 3138x - 689695000.9568112265876281

Curve labels and equations are from the PARI elldata database (Cremona tables). Ranks and L-values were computed here by PARI; no rank in this table is copied from a source. The L column for the rank 1 curve holds L'(E, 1).

04Navier-Stokes round 3: 963 rejected on measured cost, longer time instead

RUN November 20-21, 2026. Unforced blowup search, Clay statements (A) and (B) territory, on a small grid. A numerical hunt on a 643 grid cannot prove or disprove a continuum statement, and nothing here claims otherwise. Method: pseudo-spectral, nu 0.001, dt 0.01, same protocol as rounds 1 and 2.

The 963 run was measured first and rejected on the numbers. Three timing probes, 10 steps, 20 steps, and a second 10-step probe after a VM restart: 9.286, 12.335, and 17.323 seconds per step, peak RSS about 660.7 MB. The third probe extrapolates a 1000-step run to 17322.9 seconds, 4.81 hours, and the round 2 precedent says probes underestimate the full-run cost on this machine by 1.83 times, so the true cost was likely about 8.8 hours. The full 963 run was not started.

Fallback, Kida-Pelz at 643 to final time 20.0, 2000 steps: enstrophy growth 18.86 (initial 1.125, peak 21.22), max vorticity growth 11.51 (peak 34.52), energy 0.375 to 0.03311. Stayed smooth, no blowup observed. Runtime 4431.33 seconds. The peak enstrophy 21.2164 and peak vorticity 34.5187 are identical to round 2, a deterministic retrace of the first 10 time units, then the flow decays to t = 20 (enstrophy 1.6547, vorticity 7.0357). No blowup signal out to t = 20.

Reading: doubling the time horizon added decay, not growth. The unforced case stays open.

05Navier-Stokes round 4: lower viscosity, nu 0.0005

RUN November 20, 2026. Same protocol, viscosity halved to 0.0005, 643, dt 0.01, t = 10, 1000 steps. The solver was rewritten as a real-FFT variant during this round, 1.71 times faster (0.475 versus 0.815 CPU seconds per step), verified against the reference solver to near machine precision before the production runs.

Case 1, Kida-Pelz at 643, nu 0.0005: enstrophy growth 40.33 (initial 1.125, peak 45.37, final 32.68), max vorticity growth 14.65 (peak 43.96), energy 0.375 to 0.161. Stayed smooth, no blowup observed. Runtime 2008.9 seconds.

Case 2, Taylor-Green at 643 for the first time, nu 0.0005: enstrophy growth 36.42 (initial 0.375, peak 13.66, final 13.29), max vorticity growth 12.26 (peak 24.52). Stayed smooth, no blowup observed. Runtime 1421.2 seconds.

Reading: halving the viscosity more than doubled enstrophy growth against round 2 (18.86 to 40.33 for Kida-Pelz), but growth peaked and then decayed in both cases. A transient, not a runaway. The unforced case stays open.

06P versus NP round 3: the full fragment with defeaters is polynomial

RUN November 20, 2026. Repo NeuruhAI/lead-recipe-os, HEAD 7363dcb1, fetched read-only. Round 2 decided the defeater-free fragment in polynomial time and left the full fragment, with defeaters, as the open attack surface. Round 3 attacks it.

The defeater rule makes the decision easier, not harder. The dichotomy was verified exhaustively: 8,876 checks across 14 timelines, 317 signals, two floor settings, 0 violations. For every signal and every timeline, exactly one of leaf(s) and NOT(defeater leaf(s)) is SATISFIED. A SATISFIED node therefore always exists, and node-wise greedy search decides FULL-COMP in polynomial time: a valid witness was built on 14 of 14 timelines, candidate pool 3170 to 3867, 0.33 to 1.82 seconds per run. Brute-force agreement: 47 of 47 on the full fragment, 44 of 44 replicating the round 2 defeater-free check.

Hardness direction, fails on all three defeater encodings: a formula tree can distinguish SAT from UNSAT on a fixed assignment, but the search never has to use it, because a trivially SATISFIED node always exists elsewhere. One edge case is load-bearing: a hand-built row with signal_logic None and a required signal validates and evaluates SATISFIED on every timeline, but it is a META recipe, excluded from FULL-COMP by definition. The exclusion is what keeps the fragment honest.

Verdict: the full fragment admits an explicit polynomial-time search algorithm, validated against brute force. Nothing here separates or collapses P and NP, and none is claimed. The hardness question for this object is structurally impossible in the direction attempted.

07P versus NP round 4: the phase transition does not survive defeaters

RUN November 20, 2026. Repo NeuruhAI/lead-recipe-os, HEAD 7363dcb1, fetched read-only. Round 1 measured a sharp phase transition in the registry: match fraction tracking pr against evidence density. Round 4 measures it again in the full fragment with defeaters, 200 trials per cell.

Experiment A, 140 cells: plain-only recipes track pr; defeater-only recipes track (1-p)r, the mirror image, easiest to match at low density; a 50/50 mixed recipe tracks 0.5r at every density, so density stops mattering at all; a 25 percent defeater mix tracks (0.25 + 0.5p)r. Largest deviation from prediction across all 140 cells: 0.065. Experiment A2, evidential absence, 56 cells: NOT(plain) and the ABSENT operator track (0.5(1-p))r; NOT(defeater) tracks the full (1-p)r. Largest deviation: 0.055.

Experiment B, existence against density: no rare regime exists. A SATISFIED composed recipe exists on 1400 of 1400 trials at every density from p = 0.0 to 1.0, found in 1.0 to 2.0 node checks per trial. The witness kind flips with density: NOT(defeater) at p = 0, plain loose leaves at p = 1.0.

Greedy correctness argument: a proof sketch for DF-COMP-SEARCH is written up in data/p_vs_np_greedy_proof_r4.md. The theorem is that DF-COMP-SEARCH is decidable in polynomial time by one pass over a candidate node pool of at most 2,599 nodes, because ALL is SATISFIED exactly when every child is, each child evaluates independently, every composer node form is covered by the pool with dominating parameters, and a witness is then built with the real composer and validator. The scope is stated honestly: the lemmas prove the evaluator question exactly; the validity-aware answer is sound by construction and complete on all 9 battery timelines; it does not extend to arbitrary LogicNode trees or to the full fragment, where existence degenerates to always YES. One round 2 overclaim is corrected here: loose-floor dominance is not a theorem of best(), because the pool tests both floor settings.

Verdict: matching becomes rare only for fixed random recipes with several plain clauses at low density, or several defeater clauses at high density. Nothing here separates or collapses P and NP, and none is claimed.

08Yang-Mills round 3: fourth beta point and a four-point extrapolation

RUN November 20-21, 2026. Lattice 84, SU(2), Metropolis checkerboard, eps 0.35, cold start, 400 thermalization plus 1500 measurement sweeps, seed 20261370, same protocol as rounds 1 and 2. Beta 2.5, plus a 124 run at beta 2.3 for finite-volume control. Total runtime 4403.4 seconds. A lattice gap at finite spacing is a measurement, not a proof of the continuum mass gap.

Beta 2.5 on 84: average plaquette 0.653125 plus or minus 0.000275, acceptance 0.412. String tension sigma a2 0.20710 plus or minus 0.00129. Gap estimates: 0++ glueball 2.702 plus or minus 0.359, 2++ tensor 3.776 plus or minus 0.696, torelon 1.280 plus or minus 0.082 (6.4 percent error, the cleanest channel). Static potential V(1) 0.3516 plus or minus 0.0039. The torelon mass is nonzero at all four spacings with clear statistical significance.

Volume control, 124 at beta 2.3, 1000 measurement sweeps (not 1500; probes projected about 8 hours under load, so the sweep count was sized from a timing probe under the running load): plaquette 0.602131 plus or minus 0.000187, sigma 0.3172 plus or minus 0.0010, 0++ 2.368 plus or minus 0.188, torelon 2.166 plus or minus 0.642. Bulk observables show no finite-volume shift between 84 and 124.

Four-point continuum extrapolation, torelon channel, m/sqrt(sigma) against sigma a2: points 4.198 (beta 2.2), 3.306 (beta 2.3), 3.519 (beta 2.4), 2.812 (beta 2.5), non-monotonic. Weighted linear fit: continuum intercept 1.634 plus or minus 0.820, slope 6.192 plus or minus 3.476, chi2/dof 2.65/2. The fourth point dragged the intercept down by more than half against the three-point fit (3.473 to 1.634) and flipped the slope. What the fit does support: a linear a2 fit can now be performed with 2 degrees of freedom. What it does not support: a continuum claim. The intercept error is 50 percent, the slope error is 56 percent, and the fit is unstable to adding one point. This is an extrapolation attempt, not a continuum result.

09Yang-Mills round 4: smeared and variational glueball operators

RUN November 20-21, 2026. APE smearing of spatial links (alpha 0.5, levels 0, 2, 4 steps) with a variational GEVP over the three levels, at beta 2.3 and 2.4 on 84, same underlying protocol. Both betas were re-run from scratch after a VM restart killed the first attempt; checkpointing was added. Total runtime 652.9 seconds. Plain-operator numbers were reproduced bit-identically, which is the internal consistency check.

GEVP ground state 0++: 1.653 plus or minus 0.183 at beta 2.3 and 1.504 plus or minus 0.157 at beta 2.4. The plain m_eff(0) sits about 1.2 higher at both betas (2.9 sigma at 2.3, 3.5 sigma at 2.4): the plain operator was contaminated high by excited states, and smearing plus GEVP removes it. Variational gap over sqrt(sigma): 2.937 at 2.3 against 3.042 at 2.4, 3.5 percent apart, against plain 5.028 to 5.473, 8.8 percent apart. The variational gap agrees with the torelon within errors at both betas.

Caveats, stated plainly: the beta 2.3 STABILIZED verdict rests on a marginal m_eff(1) pass at 49.9 percent relative error; m_eff(2) is negative and unusable in every channel; GEVP eigenvectors have large components (minus 15.7, 48.6, minus 341.1 at 2.3), a near-degenerate basis. The 0++ gap estimate stabilizes across beta under the variational operator, but no continuum limit is taken here.

10Hodge round 3: the sheaf O_X is tractable

RUN November 20-21, 2026. Lean 4.13.0 plus mathlib v4.13.0, project under lean/. Round 2 assessed the sheaf-level exponential sequence as not tractable, because mathlib had no sheaf of holomorphic functions O_X to sheafify with. Round 3 overturns that assessment.

Piece formalized: O_X, built from mathlib's LocalInvariantProp sheaf machinery, the same way as the smooth sheaf in Mathlib.Geometry.Manifold.Sheaf.Smooth, using differentiableWithinAt_localInvariantProp with target model I(C, C). In Hodge/SheafOX.lean: holomorphicSheaf (O_X as a sheaf of types, sections exactly MDifferentiable functions), holomorphicSheafCommRing (O_X as a sheaf of commutative rings), MDifferentiableAt.inv and MDifferentiable.inv (the inverse of a nowhere-zero holomorphic function is holomorphic), holomorphic_isUnit_iff, holomorphicSheafUnits (O_X^* as a sheaf of abelian groups), mdifferentiable_exp and holomorphic_exp, holomorphicExpSheafHom (exp as a sheaf morphism O_X to O_X^*). 13 theorems in total.

Build: lake build exit 0, 3842/3842 targets, 0 sorries in project files. Axiom audit: all 13 declarations depend only on [propext, Classical.choice, Quot.sound], no sorryAx. The (1,1) case is now conditional on the sheaf long exact sequence rather than on O_X existing. Still blocked: the sheaf exponential sequence as sheaves (the inclusion Z to O_X is not built; exactness at O_X^* needs local holomorphic logarithms), the long exact sequence with c1, Dolbeault, and the Hodge decomposition. Two mathlib facts are recorded for the next round: Opens morphisms are ULift (PLift (U ≤ V)), use CategoryTheory.homOfLE; mul_inv_cancel does not apply to C (not a Group), use mul_inv_cancel₀/inv_mul_cancel₀ (GroupWithZero).

11Hodge round 4: degree-zero Dolbeault, as far as mathlib allows

RUN November 20-21, 2026. Lean 4.13.0 plus mathlib v4.13.0, project under lean/. The Dolbeault piece, formalized as far as mathlib v4.13.0 allows, and the previous agent's c1 construction recovered and verified after a VM restart killed the first attempt.

Piece formalized, in Hodge/Dolbeault.lean, two parts, 14 theorems, 4 definitions, 0 sorries. Part A, degree-zero Dolbeault on C: the Wirtinger d-bar derivative dbarDeriv f z = (fderiv R f z 1 + I * fderiv R f z I) / 2, with the theorem that a real differentiable f : C to C is complex differentiable at z exactly when its d-bar derivative vanishes. That is the kernel of d-bar equals O_X, pointwise on C. Also the global form, the forward and converse directions separately, the Cauchy-Riemann characterization, and the df/dz plus df/dzbar split. Proved from mathlib's differentiableAt_iff_restrictScalars. Part B, the type decomposition of the differential: for L : E toL[R] C on a normed space E over C, holPart is C-linear, dbarPart is C-antilinear, they sum to L, and dbarPart vanishes exactly when L is C-linear. The bridge: dbarDeriv is the dbarPart of the real differential, and complex differentiability holds exactly when the (0,1) part of the real differential vanishes.

Also recovered and verified in this round: the c1 construction in Hodge/ChernClass.lean, 9 theorems, exact at both joints in an abstract abelian category, surjecting onto the kernel, plus the sheaf-category version with c1 : H1 to H2 surjecting onto the kernel, and Lefschetz (1,1) from the sheaf long exact sequence, conditional on the Dolbeault isomorphism and the Hodge decomposition.

Build: lake build Hodge.Dolbeault exit 0, 1497/1497 targets; full lake build exit 0, 4550 targets, after the coordinator added SheafOX and Dolbeault to the library root, so the default build now covers the round 3 and round 4 work. Axiom audit of all 23 theorems (14 new plus 9 ChernClass re-verified): no sorryAx anywhere; every declaration depends only on standard axioms [propext, Classical.choice, Quot.sound]. Where the Dolbeault isomorphism stops, verified by fresh mathlib searches: no differential forms on manifolds, no exterior derivative, no de Rham cohomology; grep for dolbeault, de rham, chern, wirtinger returns only two false positives (Chernoff bound, matcherName). Still missing: the sheaves A0,q of smooth (0,q)-forms, the d-bar operator between them, Dolbeault cohomology groups, and the fine-sheaf acyclicity identifying Hq(X, O_X) with Dolbeault cohomology. The proved pieces are the two ends the isomorphism is built from; everything in between is absent.

12Navier-Stokes round 5: lower viscosity, nu 0.00025

RUN November 21, 2026. Same protocol, viscosity halved again to 0.00025, 643, dt 0.01, t = 10, 1000 steps, on the round 4 real-FFT solver. A stability probe ran first, Kida-Pelz at nu 0.00025 to t = 1.0: finite, energy 0.375 to 0.374, enstrophy peak over initial 1.37, 28.77 seconds wall, 0.288 seconds per step, far below the cost-explosion threshold. The probe passed, so the full runs went ahead and the longer-time fallback was not taken.

Case 1, Kida-Pelz at 643, nu 0.00025: enstrophy growth 64.46 (initial 1.125, peak 72.52 at t = 8.9, final 62.75), max vorticity growth 17.04 (peak 51.13), energy 0.375 to 0.210. Stayed smooth, no blowup observed. Runtime 719.2 seconds.

Case 2, Taylor-Green at 643, nu 0.00025: enstrophy growth 62.69 (initial 0.375, peak 23.51 at t = 9.9, final 23.51), max vorticity growth 13.68 (peak 27.35). Stayed smooth, no blowup observed. Runtime 422.7 seconds.

Reading: growth rose again when the viscosity was halved, in both cases. Enstrophy growth by round for Kida-Pelz: 18.86 (nu 0.001), 40.33 (nu 0.0005), 64.46 (nu 0.00025); for Taylor-Green: 36.42 (nu 0.0005), 62.69 (nu 0.00025). The trend is consistent and monotone with decreasing viscosity, but both runs stayed finite and smooth through t = 10 with energy decaying throughout. A transient, not a runaway. The peaks arrive late, at t = 8.9 and 9.9, so longer-time behaviour at this viscosity is still untested. The unforced case stays open.

13P versus NP round 5: the wider fragment

RUN November 21, 2026. Repo NeuruhAI/lead-recipe-os, HEAD 7363dcb1, fetched read-only. Round 4 proved greedy only for the composer fragment DF-COMP and left two gaps open in its proof sketch: the induction over all fifteen LOGIC_OPS, and whether a triple of satisfied leaves could validate where all singles and sampled pairs fail. Round 5 measures both, with the real evaluator, composer, and validator throughout. Total runtime 246.62 seconds.

Experiment 1, reduction check, 10 timelines: the six non-composer ops (COUNT_AT_LEAST, CONVERGENCE, CONTRADICTION_FREE, CHANGE, ACCELERATION, PERSISTENCE) appear in 0 of 558 registry recipes; they exist in LOGIC_OPS and the evaluator only. COUNT_AT_LEAST_2: 280 tested, 74 satisfied, 74 reduce to satisfied plain leaves, 0 violations. CONVERGENCE: the same counts. COUNT_AT_LEAST_3: 560 tested, 127 satisfied, 127 reduce, 0 violations. CHANGE: 11 tested, 1 satisfied, reduces. ACCELERATION: 11 tested, 6 satisfied, all reduce. PERSISTENCE: 11 tested, 1 satisfied, reduces. Greedy extends to five of the six non-composer ops by exact reduction.

CONTRADICTION_FREE is the exception, and it fails in the trivializing direction, not the hardening one. The zero-arg node is SATISFIED on every timeline including the empty one: 10 tested, 10 satisfied, 10 reduction violations. The one-arg form: 3170 tested, 3168 satisfied, 2664 violations. Admitting it makes existence always YES, the same degeneracy as defeaters but for a different reason.

Experiment 2, brute-force agreement on the extended pool (8-signal vocabulary, ALL trees over subsets of size 1 to 3): without CONTRADICTION_FREE, 44 of 44 agreement, pool 144 to 163 nodes, 174.67 seconds. With CONTRADICTION_FREE, also 44 of 44, in 1.32 seconds, because the zero-arg node makes every trial YES at size 1. Experiment 3, triple-validity rescue, exhaustive singles, pairs, and triples of satisfied leaves through the real composer and validator: all-present-8 timeline 8 of 16 valid at size 1, 72 of 112 at size 2, 272 of 448 at size 3; mixed-8 timeline 9 of 15, 82 of 90, 276 of 276; blocked-heavy timeline 0 of 6, 0 of 12, 0 of 8. Rescue found: no, in 983 combinations tried. Failure modes observed: source feasibility (monotone, all-blocked stays blocked at every size) and contradiction detection (the ownership.government NOT_GOV self-contradiction). Duplicate detection, the only plausible non-monotone rescue mechanism, never fired once.

Angles considered and closed: ingredient-deleting defeaters (no such mechanism exists in the codebase; ingredients are producer sets, never logic operands), counting and optimization (ill-posed or trivial on this codebase), and a different density law (experiment 4: perfectly correlated evidence, 20 of 20 exists in both the all-or-nothing and block-correlated cases; complementarity is per-signal and density-independent). Depth beyond 3 adds nothing. Verdict: the single-plus-pairs witness search is sufficient on every timeline tested. Nothing here separates or collapses P and NP, and none is claimed.

14Yang-Mills round 5: fifth beta point, beta 2.6

RUN November 21, 2026. Lattice 84, SU(2), Metropolis checkerboard, eps 0.35, cold start, 400 thermalization plus 1500 measurement sweeps, seed 20261380, same protocol as rounds 1 to 4, run function byte-identical to round 4. Total runtime 730.0 seconds. A lattice gap at finite spacing is a measurement, not a proof of the continuum mass gap.

Beta 2.6 on 84: average plaquette 0.670483 plus or minus 0.000240, acceptance 0.398. String tension sigma a2 0.18470 plus or minus 0.00105. Gap estimates: plain 0++ 2.714 plus or minus 0.271 (m_eff(1) not positive, so m_eff(0) is the pick), APE level 2 m_eff(0) 1.641 plus or minus 0.168, APE level 4 m_eff(0) 1.360 plus or minus 0.110, variational GEVP ground state 1.332 plus or minus 0.099, verdict STABILIZED, 2++ tensor 3.637 plus or minus 0.625, torelon 1.608 plus or minus 0.161 (picked m_eff(1)). Variational gap over sqrt(sigma): 3.099 at 2.6, against 2.937 at 2.3 and 3.042 at 2.4 in round 4, roughly flat across beta.

Five-point continuum extrapolation, torelon channel, m/sqrt(sigma) against sigma a2: points 4.198 (beta 2.2), 3.306 (beta 2.3), 3.519 (beta 2.4), 2.812 (beta 2.5), 3.742 (beta 2.6), non-monotonic. Weighted linear fit: continuum intercept 2.423 plus or minus 0.739, slope 3.236 plus or minus 3.213, chi2/dof 7.60/3. The five-point fit supports nothing firm, and it is weaker than the four-point fit, not stronger. The new 2.6 point sits above the 2.5 point, breaking the downward trend, about 2.2 combined sigma in the wrong direction for a smooth approach to a continuum limit. It pulled the intercept up from 1.634 to 2.423 and worsened chi2/dof from 1.32 to 2.53. The torelon channel itself is noisy: at 2.6 the effective masses fall steeply with tau, so the picked m_eff(1) carries real excited-state risk. Same standing caveats as round 3: sigma a2 from chi(2,2) overestimates the asymptotic tension, the 84 volume leaves finite-volume effects uncontrolled, and the linear a2 form is assumed, not established. This is an extrapolation attempt, not a continuum result. Still open: a second seed at 2.5 or 2.6 would tell whether the 2.6 point is statistics or a real break in the trend, and the variational 0++ ratio deserves its own series.

15Hodge round 5: the sheaf long exact sequence with c1

RUN November 21, 2026. Lean 4.13.0 plus mathlib v4.13.0, project under lean/. Round 4 left the sheaf long exact sequence with c1 as the next piece, conditional on the Dolbeault isomorphism and the Hodge decomposition. Round 5 proves the sequence itself for the actual exponential sequence.

Piece formalized, in Hodge/LongExactSequence.lean, 18 theorems, 0 sorries. The exponential sequence is short exact (expSeq_shortExact): 0 to Z to O_X to O_X^* to 0 as sheaves of abelian groups on the opens site. Exactness at O_X comes from the two analytic inputs: exp f = 1 implies f is locally constant in 2 pi i Z (continuity plus discreteness of Z), and every nowhere-zero holomorphic section has a local logarithm (differentiableAt_log on the slit plane). The kernel is proved categorically, and the exponential map is an epimorphism by local surjectivity. c1 is the connecting homomorphism (expSeq_chernClass), exact at both joints and surjecting onto the kernel of H2(Z) to H2(O_X). The Lefschetz bridge (lefschetz_one_one_of_expSeq) now runs with the abstract sequence replaced by the exponential sequence, conditional only on the Dolbeault isomorphism and the Hodge decomposition. Byproduct: every abelian category is balanced.

Build: lake build exit 0, 4584/4584 targets, 0 sorries in project files. Axiom audit of all 18 declarations: depends only on [propext, Classical.choice, Quot.sound], no sorryAx. Still blocked: the Dolbeault isomorphism Hq(X, O_X) isomorphic to H0,q(X) and the Hodge decomposition, because mathlib v4.13.0 has no differential forms on manifolds, so the Dolbeault groups cannot be stated. The Lefschetz (1,1) theorem is now conditional on exactly those two inputs, with the sheaf LES proved instead of assumed. Nothing here claims the Hodge conjecture or any Millennium problem is solved.

16Navier-Stokes round 6: deep dive at nu 0.000125

RUN November 21, 2026. Same protocol, viscosity halved again to 0.000125, 643, dt 0.01, t = 10, 1000 steps, on the round 4 real-FFT solver. This round is a deep dive, not a repeat: near-critical initial data, resolution scaling, and a longer-time run, so the outcome is either a candidate blowup profile or a quantitative bound on what a singularity would need. A stability probe ran first, Kida-Pelz at nu 0.000125 to t = 1.0: finite, energy 0.375 to 0.375, enstrophy peak over initial 1.37, 35.73 seconds wall, 0.357 seconds per step, far below the cost-explosion threshold. The probe passed, so the full runs went ahead and the longer-time fallback was not taken.

Case 1, Kida-Pelz at 643, nu 0.000125: enstrophy growth 88.24 (initial 1.125, peak 99.27 at t = 9.4, final 97.28), max vorticity growth 18.52 (peak 55.55), energy 0.375 to 0.267. Stayed smooth, no blowup observed. Runtime 334.0 seconds.

Case 2, Taylor-Green at 643, nu 0.000125: enstrophy growth 87.56 (initial 0.375, peak 32.84 at t = 10.0, final 32.84), max vorticity growth 19.14 (peak 38.29). Stayed smooth, no blowup observed. Runtime 344.8 seconds.

Case 3, perturbed Taylor-Green at 643, nu 0.000125: enstrophy growth 78.67 (peak 37.47 at t = 10.0), max vorticity growth 15.29. At 483 the same case grew enstrophy 58.29, so growth rises with resolution by a factor of 1.35 from 483 to 643. Stayed smooth in both. Runtime 347.4 seconds at 643.

Cases 4 and 5, anti-parallel vortex tubes at 643, nu 0.000125: the Round 1 configuration grew enstrophy 1.02 and the tight variant (closer, thinner, harder displaced, toward Kerr's original setup) grew 1.24. Both decayed to near nothing by t = 10 (final enstrophy 0.0402 and 0.0633). The colliding-tube configuration as built does not stretch at any viscosity tested, in any round. That is a finding about this attack, not about Navier-Stokes.

Case 6, Kida-Pelz at 643, nu 0.000125, to t = 20 (2000 steps): enstrophy growth 88.24, identical to the t = 10 peak, because the peak arrives at t = 9.4 and the flow then decays to 42.43 by t = 20. Max vorticity growth 20.78. The growth is transient, not runaway. Runtime 730.4 seconds.

Cost probe, 963, 10 steps with the current solver: 1.716 seconds per step, peak RSS 379.4 MB, extrapolating a full 1000-step run to 0.48 hours. Round 3 rejected 963 at about 8.8 hours with the old reference solver; with the current solver the same grid costs about half an hour, so a full 963 run is feasible in a future round. It was not started here.

Dual run: the strongest case (Kida-Pelz to t = 20, max vorticity growth 20.78) was rerun and the case hashes match exactly, so the solver is deterministic on this machine. Reading: the viscosity ladder for Kida-Pelz enstrophy growth now reads 18.86 (nu 0.001), 40.33 (nu 0.0005), 64.46 (nu 0.00025), 88.24 (nu 0.000125). Growth rises monotonically as viscosity falls, and it also rises with resolution, but every case stayed finite and smooth, and the longer-time run shows the peak decaying. A transient, not a runaway. The unforced case stays open.

17Yang-Mills round 6: second seed overturns the round 5 points

RUN November 21, 2026. Lattice 84, SU(2), Metropolis checkerboard, eps 0.35, cold start, 400 thermalization plus 1500 measurement sweeps, seed base 20261121, same protocol as round 5. Total runtime 0.0 seconds. The question is whether the round 5 five-point extrapolation instability was seed noise. The answer is yes, at least in part.

Beta 2.5, seed 2 against seed 1: average plaquette 0.652879 plus or minus 0.000336 against 0.653125 plus or minus 0.000275, agrees within errors. String tension sigma a2 0.20917 plus or minus 0.00148 against 0.20710 plus or minus 0.00129. Torelon 0.8899 plus or minus 0.0441 against 1.2798 plus or minus 0.0822, which does not agree within combined errors. The seed 1 beta 2.5 run predates the variational analysis, so there is no seed 1 GEVP to compare; seed 2 GEVP ground state 0++ is 1.6263 plus or minus 0.1034.

Beta 2.6, seed 2 against seed 1: average plaquette 0.670153 plus or minus 0.000193 against 0.670483 plus or minus 0.000240. String tension sigma a2 0.18556 plus or minus 0.00105 against 0.18470 plus or minus 0.00105, agrees within errors. Torelon 1.1931 plus or minus 0.1059 against 1.6084 plus or minus 0.1611, which does not agree within combined errors. Variational GEVP ground state 0++ 1.6030 plus or minus 0.1343 against 1.3318 plus or minus 0.0986, which also does not agree.

The torelon difference runs through every separation, not just the picked mass: at beta 2.5 the seed 2 effective masses are 2.611, 0.890, 0.216 at tau 0, 1, 2 against seed 1 values 2.747, 1.280, 0.687; at beta 2.6 they are 2.597, 1.193, 0.498 against 2.789, 1.608, 0.722. The whole correlator decays faster in the second chain. The effective masses also fall steeply with tau in both seeds, so the torelon channel carries real excited-state risk at every spacing.

Three fits of the same five-point torelon series, m/sqrt(sigma) against sigma a2: seed 1 points (4.198, 3.306, 3.519, 2.812, 3.742) give intercept 2.423 plus or minus 0.739, chi2/dof 7.60/3; seed 2 points (4.198, 3.306, 3.519, 1.946, 2.770) give intercept -0.208 plus or minus 0.643, chi2/dof 29.71/3; the inverse-variance seed-mean series gives intercept 0.702 plus or minus 0.611, chi2/dof 35.04/3. The continuum value swings from plus 2.42 to minus 0.21 depending on which Monte Carlo chain is used, and every fit has a poor chi2. The five-point linear extrapolation cannot support any continuum claim, and the jackknife errors underestimate the true chain-to-chain uncertainty. Same standing caveats as before: sigma a2 from chi(2,2) overestimates the asymptotic tension, the 84 volume leaves finite-volume effects uncontrolled, and the linear a2 form is assumed, not established. This is an extrapolation attempt, not a continuum result.

18Navier-Stokes round 7: ninety-six cubed

RUN November 21, 2026. Full 963 run at nu 0.000125, dt 0.01, t = 10, 1000 steps, on the round 4 real-FFT solver. Before the run, the snapshot loop was validated bit-identical to the round 4 solver at 323: same initial data, same step count, all result fields identical. The 963 cost probe ran first: 1.734 seconds per step, peak RSS 390.4 MB, extrapolating a full 1000-step run to 0.48 hours, matching the measured walls below. Round 3 rejected 963 at about 8.8 hours with the old reference solver; with the current solver the same grid costs about 0.38 hours.

Case 1, Kida-Pelz at 963, nu 0.000125: enstrophy growth 145.79 (initial 1.125, peak 164.02 at t = 8.2, final 132.99), max vorticity growth 28.10 (peak 84.30), energy 0.375 to 0.206. Stayed smooth, no blowup observed. Runtime 1357.6 seconds.

Case 2, Taylor-Green at 963, nu 0.000125: enstrophy growth 138.27 (initial 0.375, peak 51.85 at t = 9.8, final 51.48), max vorticity growth 24.81 (peak 49.63), energy 0.125 to 0.088. Stayed smooth, no blowup observed. Runtime 1351.5 seconds.

Case 3, Taylor-Green at 643, nu 0.000125, to t = 20 (2000 steps): enstrophy growth 88.56, barely above the t = 10 value of 87.56, because the peak arrives at t = 11.0 (peak 33.21) and the flow then decays to 14.94 by t = 20. Max vorticity growth 19.14. Taylor-Green is transient too, it just peaks later than Kida-Pelz. Runtime 617.2 seconds.

Dual run: the strongest case (Kida-Pelz at 963, max vorticity growth 28.10) was rerun and the case hashes match exactly, so the solver is deterministic on this machine. Reading: growth rises with resolution at fixed viscosity, and it does not saturate: Kida-Pelz enstrophy growth 88.24 at 643 to 145.79 at 963, a factor of 1.65; Taylor-Green 87.56 to 138.27, a factor of 1.58. The viscosity ladder for Kida-Pelz at 643 now reads 18.86 (nu 0.001), 40.33 (nu 0.0005), 64.46 (nu 0.00025), 88.24 (nu 0.000125), and the 963 value at the same viscosity is 145.79. Every case stayed finite and smooth, and both longer-time runs show the peak decaying. Transients, not runaways. The unforced case stays open.

19Yang-Mills round 7: third seed, and the estimator itself

RUN November 21, 2026. Lattice 84, SU(2), Metropolis checkerboard, eps 0.35, cold start, 400 thermalization plus 1500 measurement sweeps, seed base 20261122, same protocol as rounds 5 and 6. Total runtime 631.0 seconds. The question is whether the round 6 torelon discrepancy is statistics or a real estimator problem. The answer is both, and the estimator problem is the larger one.

Beta 2.5, seed 3 against seeds 1 and 2: average plaquette 0.653705 plus or minus 0.000315 against 0.653125 plus or minus 0.000275 and 0.652879 plus or minus 0.000336, so the plaquette agrees across seeds to about 0.001. String tension sigma a2 0.20778 plus or minus 0.00104 against 0.20710 plus or minus 0.00129 and 0.20917 plus or minus 0.00148. Torelon 1.0795 plus or minus 0.0681 against 1.2798 plus or minus 0.0822 and 0.8899 plus or minus 0.0441, and no pair agrees within combined errors. Seed 3 lands between the other two. Variational GEVP ground state 0++ 1.3555 plus or minus 0.1034 against seed 2 1.6263 plus or minus 0.1034, which does not agree within combined errors; there is no seed 1 GEVP at beta 2.5.

Beta 2.6, seed 3 against seeds 1 and 2: average plaquette 0.670055 plus or minus 0.000216 against 0.670483 plus or minus 0.000240 and 0.670153 plus or minus 0.000193. String tension sigma a2 0.18759 plus or minus 0.00098 against 0.18470 plus or minus 0.00105 and 0.18556 plus or minus 0.00105. Torelon 1.2675 plus or minus 0.0524 against 1.6084 plus or minus 0.1611 and 1.1931 plus or minus 0.1059; only seed 2 against seed 3 agrees within combined errors. Variational GEVP ground state 0++ 1.6591 plus or minus 0.1600 against 1.3318 plus or minus 0.0986 and 1.6030 plus or minus 0.1343, which also does not agree across all three seeds.

The estimator test: torelon effective masses tau by tau, for all three seeds at both betas. At beta 2.5 the seed 3 effective masses are 2.643, 1.079, 0.702 at tau 0, 1, 2 against seed 1 values 2.747, 1.280, 0.687 and seed 2 values 2.611, 0.890, 0.215; at beta 2.6 they are 2.634, 1.268, 0.941 against seed 1 2.789, 1.608, 0.722 and seed 2 2.597, 1.193, 0.498. The effective masses fall steeply with tau in every seed, so the torelon channel carries real excited-state risk at every spacing. Within a single chain the mass moves more across tau (2.395 at beta 2.5, 2.100 at beta 2.6) than it moves across chains at fixed tau (0.390, 0.415), so the plateau choice moves the mass far more than the chain does. But the chains still disagree at fixed tau beyond their combined jackknife errors, so the statistics problem is real too. The torelon channel has no stable plateau to extract a mass from, and no seed agrees with the others at the picked tau.

The three-seed inverse-variance mean series (4.198, 3.306, 3.519, 2.198, 2.969) gives intercept 2.120 plus or minus 0.526, chi2/dof 55.35/3. The series is not even monotonic in beta, and the fit is poor. OVERTURNED STANDS: the five-point linear extrapolation still cannot support any continuum claim. Same standing caveats as before: sigma a2 from chi(2,2) overestimates the asymptotic tension, the 84 volume leaves finite-volume effects uncontrolled, and the linear a2 form is assumed, not established. This is an extrapolation attempt, not a continuum result.

20Receipt chain

66 receipts, one per computation batch plus one per BSD curve plus one chain verification after reopen, per round. Chain valid: true. Verification receipt: rcpt_0066. Ledger: data/lab.db, format: sovereign runtime SovereignStore.

ReceiptKindDetailHash
rcpt_0001COMPUTATIONRiemann zero verification: first 1000 non-trivial zeros on the critical line5e08a386da7f7cf6...
rcpt_0002COMPUTATIONBSD curve 11a1: rank and L-valuecec763f6e2d22797...
rcpt_0003COMPUTATIONBSD curve 11a2: rank and L-valueaf25887ad72d1584...
rcpt_0004COMPUTATIONBSD curve 11a3: rank and L-value77d33f4ef569d0fd...
rcpt_0005COMPUTATIONBSD curve 14a1: rank and L-value9338c66c6b1ef65d...
rcpt_0006COMPUTATIONBSD curve 14a2: rank and L-value9898c4f4373ed092...
rcpt_0007COMPUTATIONBSD curve 14a3: rank and L-value64dadeb46b9317e5...
rcpt_0008COMPUTATIONBSD curve 14a4: rank and L-valuea9e1c72c85afb4ff...
rcpt_0009COMPUTATIONBSD curve 14a5: rank and L-value8905ff53a7e0bb46...
rcpt_0010COMPUTATIONBSD curve 14a6: rank and L-valuece1b8ad41a66dc4c...
rcpt_0011COMPUTATIONBSD curve 15a1: rank and L-value464e29ea589955d5...
rcpt_0012COMPUTATIONBSD curve 15a2: rank and L-valuedd9c42ba06f26e7c...
rcpt_0013COMPUTATIONBSD curve 15a3: rank and L-value662d80bb22f4dc7a...
rcpt_0014COMPUTATIONBSD curve 15a4: rank and L-value246d843c24d97b34...
rcpt_0015COMPUTATIONBSD curve 15a5: rank and L-valuea32799b96ec5d909...
rcpt_0016COMPUTATIONBSD curve 15a6: rank and L-value9903181fb50d3404...
rcpt_0017COMPUTATIONBSD curve 15a7: rank and L-valuef0ce62dfa66524a2...
rcpt_0018COMPUTATIONBSD curve 15a8: rank and L-value592bcbc9b176dfca...
rcpt_0019COMPUTATIONBSD curve 17a1: rank and L-valuef27bf60b3dce3bd7...
rcpt_0020COMPUTATIONBSD curve 17a2: rank and L-value59b21533385a2548...
rcpt_0021COMPUTATIONBSD curve 17a3: rank and L-valuee24ee8df2a7b0d6b...
rcpt_0022COMPUTATIONBSD curve 17a4: rank and L-valuecdef991479c91de0...
rcpt_0023COMPUTATIONBSD curve 19a1: rank and L-valuee15da00aeb8e7c6c...
rcpt_0024COMPUTATIONBSD curve 19a2: rank and L-valued6a5e200d3d49b5e...
rcpt_0025COMPUTATIONBSD curve 19a3: rank and L-value21c05c1e0964dfcc...
rcpt_0026COMPUTATIONBSD curve 20a1: rank and L-value1bda09f19c9eff0b...
rcpt_0027COMPUTATIONBSD curve 20a2: rank and L-value0743d1b4bb0ca9e1...
rcpt_0028COMPUTATIONBSD curve 20a3: rank and L-valuedbbb70cbc3901569...
rcpt_0029COMPUTATIONBSD curve 20a4: rank and L-value10c0a8405f9436c5...
rcpt_0030VERIFICATIONReceipt chain verified after store reopen (restart proof)328969b5d1e89f98...
rcpt_0031COMPUTATION32^3 pseudo-spectral Taylor-Green run to t=10, nu=0.01, peak enstrophy growth 1.73, stayed smooth; hunt to t=10 at nu=0.001, strongest Kida-Pelz 48^3 enstrophy growth 18.29, max vorticity growth 7.97, stayed smooth, no blowup observede1a15151a9566ffb...
rcpt_0032AUDITBackground reading note on the OpenAI Navier-Stokes paper, read 2026-11-20, no verdict per founder correction; Clay clock not started, unforced case open, priority dispute unresolved50007e53efd59cef...
rcpt_0033AUDITP vs NP recipe-search complexity object: registry search in P by measurement (558 recipes, 0.0718 s sweep), composition search 2^317 candidates with linear verification, defeater non-monotonicity confirmed by run, phase transition measured; barriers: relativization no, natural proofs yes, algebrization no; no separation or collapse claimed39e22c544e242c44...
rcpt_0034COMPUTATION8^4 SU(2) Wilson lattice, beta 2.2 and 2.4, avg plaquette 0.568932 and 0.631483, string tension sigma a^2 0.4054 and 0.2444, torelon gap 2.67 and 1.740c03a14f69f03b5f...
rcpt_0035AUDITGap audit: Wightman construction judged proved by Nov 10 assessment, existence in Clay sense disputed by Nielsen critique, mass gap unclaimed and open, Clay clock cannot start1a4f631c8cf5f73a...
rcpt_0036FORMALIZATIONLean 4.13.0 + mathlib v4.13.0: compiled; Lefschetz (1,1) proved conditionally and kernel-checked with no sorry, general Hodge conjecture stated; unconditional case blocked by the missing exponential sequence and Hodge decomposition in mathlib94af29c9c282f1ce...
rcpt_0037VERIFICATIONReceipt chain verified after store reopen (restart proof)298ac3a407ed8312...
rcpt_0038COMPUTATIONRound 2: Kida-Pelz 64^3 enstrophy growth 18.86, max vorticity growth 11.51, stayed smooth, 1516.7 s; high-amplitude anti-parallel tubes 48^3 growth 1.02, stayed smooth; no blowup observedde27ed63b14d088a...
rcpt_0039AUDITRound 2: defeater-free fragment DF-COMP admits explicit polynomial-time greedy search, at most 2599 candidates, 0.10-0.58 s per run, 44 of 44 brute-force agreement; 3-SAT hardness direction stalls structurally; no separation or collapse claimed2e6be66bc5335930...
rcpt_0040COMPUTATIONRound 2: 8^4 SU(2) Wilson lattice beta 2.3, plaquette 0.602788, sigma a^2 0.3166, torelon gap 1.8602; three-point continuum extrapolation intercept 3.473 plus/minus 1.506, slope unconstrained; 10^4 volume check consistentf45e4199d9c5fc9d...
rcpt_0041FORMALIZATIONRound 2: Lean 4.13.0 + mathlib v4.13.0, exponential sequence 0 to Z to C to C* to 1 proved exact at every term, 9 new theorems, 0 sorries, no sorryAx; sheaf-level sequence still blocked, unconditional (1,1) still conditionalc23ff4231dd9dc6c...
rcpt_0042COMPUTATIONRound 2 calibration: BSD conductor to 30, 57 curves, all rank 0, all L(E,1) nonzero; Riemann first 2000 zeros re-verified on the line, max deviation 0.0; no pattern breaks12b0bad2229d1ad0...
rcpt_0043VERIFICATIONReceipt chain verified after store reopen (restart proof), round 22497e15506a5db7c...
rcpt_0044COMPUTATIONRound 3: 8^4 SU(2) Wilson lattice beta 2.5, plaquette 0.653125, sigma a^2 0.20710, torelon gap 1.280 plus/minus 0.082; 12^4 volume control at beta 2.3 shows no finite-volume shift; four-point torelon extrapolation intercept 1.634 plus/minus 0.820, slope flipped vs three-point, supports nothing firm797a3954cc99f469...
rcpt_0045COMPUTATIONRound 3: 96^3 REJECTED on measured timing (10-step probe 17.323 s/step, peak RSS 660.7 MB, extrapolated 4.81 h, probes underestimate by 1.83x); fallback Kida-Pelz 64^3 to t=20, 2000 steps, 4431.3 s, enstrophy peak 21.2164 then decay to 1.6547, stayed smooth, no blowup signal out to t=20601b73f2d8597401...
rcpt_0046AUDITRound 3: full fragment with defeaters is POLYNOMIAL, not hard; defeater rule makes the decision easier, a SATISFIED node exists on every timeline, greedy builds a valid witness on 14 of 14 timelines, 47 of 47 brute-force agreement; three 3-SAT reduction attempts all failed with failure modes recorded; no separation or collapse claimed7180d93c9ce7e61b...
rcpt_0047FORMALIZATIONRound 3: Lean 4.13.0 + mathlib v4.13.0, sheaf O_X IS tractable, overturning the round 2 assessment; holomorphicSheaf, holomorphicSheafCommRing, holomorphicSheafUnits, holomorphicExpSheafHom, 13 theorems, 0 sorries, no sorryAx; (1,1) case now conditional on the sheaf long exact sequence, not on O_X existing344a776c0a612a6b...
rcpt_0048COMPUTATIONRound 4: APE-smeared variational (GEVP) glueball operators at beta 2.3 and 2.4 on 8^4; GEVP ground state 0++ 1.653 plus/minus 0.183 at 2.3 and 1.504 plus/minus 0.157 at 2.4, plain m_eff(0) was contaminated high by about 1.2; variational gap over sqrt(sigma) 2.937 vs 3.042, 3.5 percent apart vs 8.8 percent plain; agrees with torelon within errors; no continuum limit taken05e5a7d360e7a2f2...
rcpt_0049COMPUTATIONRound 4: nu 0.0005, 64^3, T=10, both cases stayed smooth; Kida-Pelz enstrophy growth 40.33, max vorticity growth 14.65, 2008.9 s; Taylor-Green first time at 64^3, enstrophy growth 36.42, vorticity growth 12.26, 1421.3 s; halving viscosity more than doubled enstrophy growth vs round 2, transient not runawayed5e808dba468fea...
rcpt_0050AUDITRound 4: the p^r phase transition does NOT survive defeaters; plain tracks p^r, defeater tracks (1-p)^r, 50/50 mixed tracks 0.5^r at every density, max deviation 0.065 across 140 cells; no rare regime exists, a SATISFIED composed recipe exists on 1400 of 1400 trials; greedy correctness proof sketch written, DF-COMP-SEARCH decidable in polynomial time by one pass over at most 2599 nodese19ab5074df8e70c...
rcpt_0051FORMALIZATIONRound 4: Lean 4.13.0 + mathlib v4.13.0, degree-zero Dolbeault formalized as far as mathlib allows; dbar derivative, kernel of d-bar equals O_X pointwise on C, (1,0)/(0,1) type decomposition of the differential, 14 new theorems plus 9 ChernClass theorems re-verified, 0 sorries, no sorryAx; full Dolbeault isomorphism still blocked, no differential forms on manifolds in mathlib v4.13.02cb0569ba21a0ada...
rcpt_0052COMPUTATIONRound 4 calibration: BSD conductor to 40, 93 curves, 92 rank 0 and 1 rank 1 (37a1), all 92 rank-0 L(E,1) nonzero; Riemann first 3000 zeros re-verified on the line, max deviation 0.0, first 2000 match round 2 exactly; run twice, identical hashes660f5ee3b806be69...
rcpt_0053VERIFICATIONReceipt chain verified after store reopen (restart proof), rounds 3 and 4a29fb21ba5da38a5...
rcpt_0054COMPUTATIONRound 5: fifth beta point, 8^4 SU(2) Wilson lattice beta 2.6, plaquette 0.670483, sigma a^2 0.1847, torelon 1.608 plus/minus 0.161, GEVP 0++ 1.332 plus/minus 0.099; five-point torelon extrapolation intercept 2.423 plus/minus 0.739, chi2/dof 7.60/3, weaker than the four-point fit, supports nothing firm; variational gap over sqrt(sigma) 3.099, flat across 2.3/2.4/2.65e14340707a49559...
rcpt_0055COMPUTATIONRound 5: lower viscosity nu 0.00025, 64^3, dt 0.01, T=10, stability probe passed (28.77 s, no cost explosion); Kida-Pelz enstrophy growth 64.46, peak 72.52 at t=8.9, max vorticity growth 17.04; Taylor-Green enstrophy growth 62.69, peak 23.51 at t=9.9, max vorticity growth 13.68; both stayed smooth, growth monotone with decreasing viscosity, transient not runaway834d8e60f165a82c...
rcpt_0056AUDITRound 5: wider fragment, arbitrary defeater-free LogicNode trees over all 15 LOGIC_OPS; greedy extends to 5 of 6 non-composer ops by exact reduction (0 non-CF violations), extended brute force 44 of 44 agreement; CONTRADICTION_FREE is the exception, zero-arg node SATISFIED on every timeline, makes existence trivially always YES; no triple validates where singles and pairs fail (0 rescues in 1066 combos); ingredient-deleting defeaters and counting/optimization closed as not a mechanism / ill-posed; no separation or collapse claimed108d282502990ffd...
rcpt_0057FORMALIZATIONRound 5: Lean 4.13.0 + mathlib v4.13.0, sheaf long exact sequence WITH c1 for the actual exponential sequence, Hodge/LongExactSequence.lean, 18 theorems, 0 sorries, no sorryAx; expSeq_shortExact (0 -> Z -> O_X -> O_X^* -> 0 as sheaves), c1 exact at both joints and surjecting onto the kernel, Lefschetz (1,1) now conditional only on the Dolbeault isomorphism and the Hodge decomposition; full Dolbeault isomorphism still blocked, no differential forms in mathlib v4.13.04f990461c0909ec1...
rcpt_0058COMPUTATIONRound 5 calibration: BSD conductor to 50, 130 curves, 128 rank 0 and 2 rank 1 (37a1 and new 43a1), all rank-0 L(E,1) nonzero; Riemann first 4000 zeros re-verified on the line, max deviation 0.0, first 3000 match round 4 exactly; run twice, identical hashes8313d8883464866f...
rcpt_0059VERIFICATIONReceipt chain verified after store reopen (restart proof), round 5cb53f599378c1242...
rcpt_0060COMPUTATIONRound 6: deep dive at nu 0.000125, 64^3, dt 0.01, stability probe passed (35.73 s, per-step 0.357 s, no cost explosion); Kida-Pelz enstrophy growth 88.24 (ladder 18.86, 40.33, 64.46, 88.24), Taylor-Green 87.56, perturbed Taylor-Green 78.67 at 64^3 and 58.29 at 48^3 (resolution ratio 1.35); anti-parallel tubes inert again (1.02, and tight variant 1.24), the configuration does not stretch at any tested viscosity; Kida-Pelz T=20 peaks at t=9.4 and decays to 37.7x by t=20, growth is transient not runaway; 96^3 cost probe with the current solver 1.716 s/step, full run about 0.48 h and 379 MB peak RSS, the Round 3 8.8 h rejection was the old solver; dual run of the strongest case identical hashes; all cases stayed smooth, no blowup4d0929c10c679f68...
rcpt_0061COMPUTATIONRound 6: second seed (base 20261121) at beta 2.5 and 2.6, 8^4 SU(2), same setup as round 5; seed 2 torelon is far below seed 1 at both betas (2.5: 0.890 plus/minus 0.044 vs 1.280 plus/minus 0.082; 2.6: 1.193 plus/minus 0.106 vs 1.608 plus/minus 0.161), and the difference runs through every tau, so it is a correlator-level chain difference, not plateau picking; GEVP 0++ seed 2 is higher at 2.6 (1.603 plus/minus 0.134 vs 1.332 plus/minus 0.099); torelon extrapolation OVERTURNED: seed 1 fit intercept 2.423 plus/minus 0.739 chi2/dof 7.60/3, seed 2 fit intercept -0.208 plus/minus 0.643 chi2/dof 29.71/3, seed-mean fit 0.702 plus/minus 0.611 chi2/dof 35.04/3; the five-point linear extrapolation cannot support any continuum claim, and the jackknife errors underestimate the true chain-to-chain uncertainty87222c6ab1e1234b...
rcpt_0062VERIFICATIONReceipt chain verified after store reopen (restart proof), round 64ec224b6ae44993c...
rcpt_0063COMPUTATIONRound 7: full 96^3 Navier-Stokes at nu 0.000125, dt 0.01, t = 10, on the round 4 real-FFT solver; snapshot loop validated bit-identical to the round 4 solver at 32^3 (all result fields identical); Kida-Pelz 96^3 enstrophy growth 145.79 (peak 164.02 at t = 8.2, initial 1.125, final 132.99), max vorticity growth 28.10 (peak 84.30), energy 0.375 to 0.206, wall 1357.8 s; Taylor-Green 96^3 enstrophy growth 138.27 (peak 51.85 at t = 9.8, final 51.48), max vorticity growth 24.81 (peak 49.63), energy 0.125 to 0.088, wall 1351.7 s; growth rises with resolution at fixed viscosity: Kida-Pelz 88.24 at 64^3 to 145.79 at 96^3 (ratio 1.65), Taylor-Green 87.56 to 138.27 (ratio 1.58), no saturation; Taylor-Green 64^3 to T = 20 enstrophy growth 88.56 (peak 33.21 at t = 11.0, final 14.94), max vorticity growth 19.14, wall 617.3 s, so Taylor-Green is transient too, it just peaks later than Kida-Pelz; 96^3 cost probe 1.734 s/step, peak RSS 390.4 MB, full 1000-step run about 0.38 h, matching the measured walls; dual run of the strongest case (Kida-Pelz 96^3) identical hashes; all cases stayed smooth, no blowupdc84f23b7fdf0901...
rcpt_0064COMPUTATIONRound 7: third seed (base 20261122) at beta 2.5 and 2.6, 8^4 SU(2), same setup as rounds 5-6, runtime 631.0 s; seed 3 torelon lands between seeds 1 and 2 at both betas (2.5: 1.080 plus/minus 0.068 vs seed 1 1.280 plus/minus 0.082 and seed 2 0.890 plus/minus 0.044, no pair agrees within combined errors; 2.6: 1.268 plus/minus 0.052 vs 1.608 plus/minus 0.161 and 1.193 plus/minus 0.106, only seed 2 vs seed 3 agree); GEVP 0++ seed 3 is 1.356 plus/minus 0.103 at 2.5 (seed 2 1.626, no seed 1) and 1.659 plus/minus 0.160 at 2.6 (seed 1 1.332 plus/minus 0.099, seed 2 1.603 plus/minus 0.134); estimator test at both betas: within a single chain the torelon effective mass moves more across tau (2.395 at 2.5, 2.100 at 2.6) than it moves across chains at fixed tau (0.390, 0.415), so the correlator has no stable plateau and the plateau choice moves the mass far more than the chain does, while the chains still disagree at fixed tau beyond combined errors; the three-seed inverse-variance mean series fit gives intercept 2.120 plus/minus 0.526, chi2/dof 55.35/3; OVERTURNED STANDS: the five-point linear extrapolation still cannot support any continuum claim6dac4cdca5d8baff...
rcpt_0065COMPUTATIONField renders for rounds 5, 6 and 7, 1080x1350, computed from the lab's own solvers, same pattern as rounds 1-4. Round 5: Kida-Pelz 64^3 nu 0.00025 enstrophy growth 64.46 (peak t = 8.9), Yang-Mills weave beta 2.6 plaquette 0.673439, PNG millennium-field-render-r5.png 390337 bytes sha256 f724de7b6c36dff9.... Round 6: Kida-Pelz 64^3 nu 0.000125 enstrophy growth 88.24 (peak t = 9.4), Yang-Mills weave beta 2.5 plaquette 0.634856, PNG millennium-field-render-r6.png 387365 bytes sha256 e8fc1bc5eabddef2.... Round 7: Kida-Pelz 96^3 nu 0.000125 enstrophy growth 145.79 (peak t = 8.2), Yang-Mills weave beta 2.6 plaquette 0.650786, PNG millennium-field-render-r7.png 422766 bytes sha256 615be1de2606df1d....8b803e24002018d7...
rcpt_0066VERIFICATIONReceipt chain verified after store reopen (restart proof), round 7bdebd1e7d93dd108...

21How to re-run

Everything is in ~/workspace/millennium-lab/. Round 7: Navier-Stokes: python3 compute/navier_stokes_hunt_r7.py (includes the 963 cost probe, the snapshot-loop validation, the longer-time Taylor-Green, and the dual run of the strongest case). Yang-Mills: python3 compute/wilson_lattice_r7.py (third seed, base 20261122, at beta 2.5 and 2.6; checkpoints after each beta; includes the tau-by-tau estimator test and the three-seed mean fit). Field renders: python3 poster/render_fields_r567.py 5, then 6, then 7, then python3 poster/compose_r567.py 5, 6, 7. Receipts: python3 compute/receipts_append_r7.py (opens data/lab.db WITHOUT deleting it; writes data/lab.db and data/receipts.json; resume-safe if a partial run already appended the Navier-Stokes and Yang-Mills receipts). Page: python3 compute/build_page.py (CSS at compute/atlas-style.css). Never run compute/receipts.py on the live ledger: it deletes the ledger and rebuilds runs 1-2 only.

Round 6: Navier-Stokes: python3 compute/navier_stokes_hunt_r6.py (includes the stability probe, the 963 cost probe, and the dual run of the strongest case; the longer-time fallback was not taken). Yang-Mills: python3 compute/wilson_lattice_r6.py (second seed, base 20261121, at beta 2.5 and 2.6; checkpoints after each beta). Receipts: python3 compute/receipts_append_r6.py (opens data/lab.db WITHOUT deleting it; writes data/lab.db and data/receipts.json). Page: python3 compute/build_page.py (CSS at compute/atlas-style.css). Never run compute/receipts.py on the live ledger: it deletes the ledger and rebuilds runs 1-2 only.

Round 5: Riemann: ./pari/bin/gp -q < compute/riemann_scan_r5_upper.gp for the new interval [3545, 4650] (the full [0, 4650] scan overflows the default 8 MB PARI stack; split the interval), then python3 compute/riemann_verify_r5.py (mpmath 1.2.1, system Python 3.12.3). BSD: ./pari/bin/gp -q < compute/bsd_curves_r5.gp, then python3 compute/bsd_curves_r5.py. Navier-Stokes round 5: python3 compute/navier_stokes_hunt_r5.py (includes the stability probe; the longer-time fallback was not taken). P vs NP round 5: python3 compute/p_vs_np_r5.py (fetches NeuruhAI/lead-recipe-os read-only). Yang-Mills round 5: python3 compute/wilson_lattice_r5.py. Hodge: cd lean && lake build (Lean 4.13.0, mathlib v4.13.0; elan at ~/.elan/bin). Receipts: python3 compute/receipts_append_r5.py (opens data/lab.db WITHOUT deleting it; writes data/lab.db and data/receipts.json).

Earlier rounds: Riemann round 4: ./pari/bin/gp -q < compute/riemann_scan_r4.gp, then python3 compute/riemann_verify_r4.py. BSD round 4: ./pari/bin/gp -q < compute/bsd_curves_r4.gp, then python3 compute/bsd_curves_r4.py. Navier-Stokes round 3: python3 compute/navier_stokes_hunt_r3.py (includes the 963 timing probe; the full 963 run was not started). Navier-Stokes round 4: python3 compute/navier_stokes_hunt_r4.py. P vs NP round 3: python3 compute/p_vs_np_r3.py; round 4: python3 compute/p_vs_np_r4.py (both fetch NeuruhAI/lead-recipe-os read-only). Yang-Mills round 3: python3 compute/wilson_lattice_r3.py; round 4: python3 compute/wilson_lattice_r4.py.

22Sources and audit

Atlas: The Six Unsolved, which names the six first moves this lab executes. Problem statements and prize rules: Clay Mathematics Institute. Riemann verification background: the first 1013 zeros verified, about 41 percent of zeros proved on the line (atlas sources). PARI/GP 2.19.1 source and elldata from pari.math.u-bordeaux.fr. mpmath 1.2.1. Lead Recipe OS: NeuruhAI/lead-recipe-os, HEAD 7363dcb1, read 2026-11-21. Every number on this page comes from a run on November 20 or 21, 2026. Audit date: November 21, 2026.

Neuruh, AI That Executes. Contact: jeramie@neuruh.com · Scheduling: cal.com/neuruh
Every number on this page carries its run date. A stat without a date is a rumor with better formatting.
Millennium Lab v6.0, 2026-11-21. Three moves run, round 7 complete. Companion page: The Six Unsolved.