L5 M2 batch (written before launch). CP-SAT 1 worker, 180 s, seed 1, real vs rewirings r1-r3.
Families: radon (g=-1, side 2p) p=7,11,19,23,31,43 mode mixed; two (xy=1,-1, side 3p) p=11,19,23,31 mode mixed; one (xy=1, side 2p, HJSW control) p=11,19,31,43 mode pure.
Readings (from MR172): within 1% of side at every p for radon and two -> M supported (55%); real >= rewired + 2% side, gap not shrinking -> M refuted (15%); real below rewired -> algebra hurts (20%, raised by E166 precedent); else inconclusive.
Control: 'one' pure rewiring must differ from real by >= 2% of side, else the method cannot see fibre structure and the main reading is void.
Only OPTIMAL-vs-OPTIMAL pairs count as exact; others are matched-budget values.

ADDENDUM (written after the first batch, before the control rerun): the 'one' control in batch 1 was VOID. At phase 0 a single hyperbola at side 2p has optimum about n (only 4-point lines; 22 at p11), and it has no pure 3-point lines (target 0). The phase ((p+1)/2, p-1) attains 3(p-1) exactly at p = 11-43. Control 2: mode 'all' (every rich line rewired, sizes and per-point degrees kept) at that phase, p = 11, 19, 31, 43, 180 s. Pass: real exceeds the rewired mean by >= 2% of side at every p. The 'two' family ran at phase 0 and may be degenerate in the same way; its rows are provisional.

CONTROL 2 RESULT (22:04Z): PASS. One hyperbola at phase ((p+1)/2, p-1), all lines rewired: real 30/54/90/126 = 3(p-1), all OPTIMAL, against rewired means 27.7/49.3/81.3/111.3 (all OPTIMAL). Gap/side .106/.123/.140/.171, growing. The method sees algebraic structure when it exists.
BATCH 3 (before launch): 'twos' = two levels xy = 1, -1 at the same phase, side 3p, mode mixed, p = 11, 19, 23, 31, 120 s. CPU note: batches 1-2 used up to 3.6 CPU-h against the prestated 3; batch 3 adds <= 0.5 CPU-h because control 1 was void. No longer radon runs.
BATCH 3 RESULT (22:15Z): twos gap/side -.040/-.023/-.039/-.029 at p = 11/19/23/31; none OPTIMAL. Real below rewired at every p.
CONFOUND CHECK (before launch): matched-budget values could penalise the real instance if it is harder for CP-SAT. Check:
twos p=7 (side 21) real, r1-r3, 300 s; if all OPTIMAL this is a second exact pair. Reading: real < rewired mean at OPTIMAL
-> the deficit is not a solver artefact at that p; real = rewired -> the matched-budget deficit may be solver difficulty,
and the reading drops to "inconclusive" for p >= 11.
CONFOUND 2 (22:20Z, before any linear run): rewired instances are NOT linear. Real line systems cover each point pair at most
once; rewired ones cover 216-228 pairs twice (radon p19) and 382-394 (twos p11). Triples sharing a pair are easier to satisfy
together, so nonlinearity can favour the rewired side and may explain "real below rewired". Fix: linear rewiring (swap repair
until no pair is covered twice by keep + target, degrees and component patterns kept). Rerun radon p7-31 and twos p7-31 at the
same budgets. Readings: linear gap still <= -2% side at most p -> "algebra hurts" stands; gap within 1% -> M supported;
gap >= +2% -> M refuted. Batches 1 and 3 are superseded by the linear rerun for the M question.
P7 CHECK RESULT (22:23Z): twos p7 real 33 (bound 36, not OPTIMAL at 300 s); r1-r3 34 OPTIMAL. Not an exact pair: the real
instance is harder for CP-SAT, so matched-budget gaps are partly solver difficulty. Rerun: real p7, 4 workers, 1200 s.
Linear repair v1 (2M swaps, 45 s) leaves 36-77 doubly covered pairs; v2 needed before the linear rerun.
LINEAR REWIRING v2 (22:31Z): repair moves only points of doubly covered pairs, annealed. 0 doubly covered pairs, degrees
and patterns exact, < 2 s, at radon 7/19/43, twos 7/11/31, one 43. Variants l1-l3 (rng seeds 1-3).
LINEAR RERUN (before launch): radon p7-43 l1-l3 at 180 s; twos p7-31 l1-l3 at 120 s (twos p7 also 300 s); one p11-43
mode all l1-l3 at 180 s (control, should still PASS). Compare with the existing real rows at the same budget.
Exact pairs (both OPTIMAL) are reported separately and decide the sign where they exist.
DECIDE (22:52Z, before launch): twos p7 real stayed 33/36 at 1200 s x 4 workers. Decide size >= 34 as feasibility, 2 workers, 900 s. INFEASIBLE -> real optimum 33 < 34 = rewired (exact pair, gap -.048). FEASIBLE -> real >= 34, matched-budget deficit was solver difficulty.
DECIDE RESULT (23:05Z): twos p7 real, size >= 34: UNKNOWN at 900 s x 2 workers. No exact pair for twos at p7. The real instance is much harder to prove than its rewirings (they reach OPTIMAL 34 in < 300 s, 1 worker); a plausible cause is its large symmetry group. NEXT (before launch): twos p5 real, r1-r3, l1-l3, 300 s, as the smallest possible exact pair.
EXACT SEEDS (23:06Z, before launch): twos p5 exact: real 21, r 22/22/22, l 21/22/21. To get the exact distribution, run l4-l12 at radon p7 and twos p5 (300 s, 2 threads). Report fraction of linear rewirings with optimum > real, = real, < real.
EXACT SEEDS RESULT (23:07Z): all OPTIMAL. radon p7: real 23; linear l1-l12: 7 equal 23, 5 reach 24, none below (mean 23.42, gap -.030 side). twos p5: real 21; l1-l12: 7 equal, 5 reach 22, none below (mean 21.42, gap -.028). Nonlinear r1-r3 reach 24 and 22 every time. Real is at the minimum of the linear distribution at both exact sizes. Note: other agents' processes share this host; my runs used <= 4 threads.
LINEAR RERUN RESULT (23:34Z): radon gap (linear) 0/0/+.036/-.022/-.035 at p = 11/19/23/31/43 (sign unstable). twos -.016/-.040/-.023/-.019/-.032 at p = 7/11/19/23/31: negative at every p, matched budget. twos p7 l1,l2 also stop at 33 (not OPTIMAL, 300 s); only l3 is OPTIMAL 34, so part of the real instance's solver hardness is shared by linear rewirings. Reading for twos: 'real below rewired' at matched budget, not established as optima.
SYMMETRY PROBE (23:41Z): CP-SAT symmetry_level 4 on twos p7: real has 1 generator (an involution, 54 orbits of size 2); linear l1 has none. Both are 33/36 after 60 s. So the hardness is shared with linear rewirings and is not a symmetry effect; the easy instances were the nonlinear r-rewirings. Correction: pitfall 2 in N44jL8Cn overstates the real instance's handicap against linear rewirings.
SINGLE-COMPONENT CHECK (00:22Z, before launch): one component alone is a trivial lower bound on the union optimum. If a single component's CP-SAT value exceeds the union's matched-budget value at the same p, the union rows at that p are below a trivial bound and say nothing about optima. Runs: radon c0,c1 p=7,11,19,43; twos c0,c1 p=5,11,31; 120 s.
SINGLE RESULT (00:36Z): radon one parabola = 15/23/39/87 OPTIMAL at p7/11/19/43 (1.07-1.01 side); union real 23/35/56/112, so radon rows are above the trivial bound. twos one level = 19/44 OPTIMAL at p5/11; p31: 124 (c1) and 119 (c0), FEASIBLE. Union values at twos p31 (real 115, rewired 116-120) are BELOW the single-level value 124: the prestated trigger fires, twos p31 rows say nothing about optima. Running twos p19/p23 singles to see which other rows are affected.
SINGLE RESULT 2 (00:42Z): twos one level at p19 = 76/76 (both comps), p23 = 93/92, all FEASIBLE at 120 s, = 4p (1.33 side), as at p11 (44 OPT). Union at 120 s: p19 real 78, rewired 78-81; p23 real 95, rewired 95-98; p31 below (115-120 < 124). So at p >= 19 every twos union row is within 0-5 points of the one-level floor 4p, which the union must contain. Reading: twos rows at p >= 19 measure the solver near the trivial floor, not optima; the "real lower every time" pattern is not evidence for or against M. Valid twos evidence: exact p5, and p7/p11 only weakly. Radon rows stand (well above floor).
SPLIT PROBE (00:50Z, before launch): per-component split of the real union selection, radon p7/11/19/43, twos p5/7/11, 90 s. Input to MR173 conjecture X (exclusive density). Reading: minority share falling with p supports X; minority share flat near 1/2 at all p is evidence against X at finite p.

CENSUS BIG (01:40Z 11-16): radon D_mixed 23.69/26.85/31.8 at p131/199/307 (D_pure3 ~0.14). D/ln p 2.4 -> 5.5 (p7..307); D/ln^2 p 1.23 -> 0.97.
MR173 LAW K (written 01:50Z, before any L6 run): f = OPT/|U| = c*sqrt(ln D / D). Fit from radon best-found: c = 0.715/0.734/0.772 at p7/19/43.
 With c = 0.77: predicted union gain over one parabola (2p+1), as fraction of side: p131 0.12, p199 0.08, p307 0.01. Crossover D ~ 33-37, p ~ 350-500.
 T1 (p19, scaled linear rewiring, D in 5/9/15/25/40): predicted f(D)/sqrt(lnD/D) flat within +-15%; union reaches single (39-41) by D = 25-40.

L6 M0 FINDING (~02:12Z 11-16 [timestamp corrected; I first wrote 02:25Z by guess], after the 01:50Z prestatement): Paper 99 Table 2 already has LNS caps for the radon pair:
 35/57/121/179/219 at p11/19/43/67/83 (g=-1 at p11-43; g=2 at p67/83, D_mixed 18.99/20.8 by my census).
 My CP-SAT best at p43 was 112, so my fit used weak values. With P99 values: c = 0.715/0.76/0.833/0.848/0.863 (p7/19/43/67/83), RISING.
 Gain over 2p+1 per side: 0.571/0.545/0.474/0.395/0.328/0.313 (p7..83). The c=0.77 law predicted ~0.17 at p83; actual >= 0.31.
 Reading: the law's FORM is wrong at fixed c (my "form wrong, direction right" branch, provisionally); direction (gain falling) holds.
 Gain vs ln p is roughly linear, slope ~ -0.10, which would reach 0 near ln p ~ 7.5 (p ~ 1800). Not a prediction yet.
 Change to T2: use radon's LNS (P99 supplement src/export_c1c2_cap_witnesses.cpp) as the solver, not CP-SAT. Single-parabola value at p67/83 must be checked (only verified 2p+1 to p43).

L6 T2 PRESTATED (~02:16Z 11-16 [corrected from a guessed 02:35Z], before T2 runs, launched 02:17Z). Solver: code/l6/lns (generic port of P99's LNS; checks every line <= 2 at exit).
 Calibration (already run): radon g=-1 p43 union 125 in 20 s (2 seeds), above P99's 121 and my CP-SAT 112; one parabola LNS 86 (exact 87).
 So my CP-SAT union values were weak by ~0.15 side at p43. Gains below use the best LNS union minus the single-parabola
 UPPER bound (CP-SAT bound, = value if OPTIMAL) -> conservative toward K.
 Runs: union p67/83 (300 s), p131 (600 s), p199 (900 s), 3 seeds each, g=-1 where p=3 mod 4 and chi(-1)=-1 (all p=3 mod 4).
 Single comp0 CP-SAT 300 s at p67/83/131/199. Total ~2.1 CPU-h, 4 parallel.
 Thresholds unchanged from MR173: refuted if gain >= 0.25 side at p199 or >= 0.35 at p131; supported if <= 0.12 at p199.
 My expectation now, after M0 (P99 table: 0.395/0.328/0.313 at p43/67/83 with 1-s LNS): refuted-at-threshold 60%,
 form-wrong band 35%, supported 5%. A threshold refutation refutes the fitted rate, not "gain -> 0" itself.

L6 T1 PRESTATED (~02:24Z 11-16, before T1 runs). DEVIATIONS from MR173 plan, stated now:
 (a) p19 with LNS (60 s, 2 seeds) instead of p11 CP-SAT 600 s (p11 CP-SAT was not OPTIMAL at 180 s; LNS is the stronger tool).
     p7 keeps CP-SAT exact at D = 5/9/15/25 (300 s).
 (b) Random mixed triples are uniform 2+1 (component chosen 50/50), linear w.r.t. kept lines and each other, NOT degree-matched
     per point (Poisson-like degrees). Real D at p19 = 9.06; the D9 instance is the matched-mean control.
 Prediction kept from MR173: f/sqrt(lnD/D) flat within +-15% across D; union -> single (39 at p19, 15 at p7) by D = 25-40.
 Also run: real p19 union with LNS 60 s (reference for D9).
 T1 note (02:20Z): at p7 linear random 2+1 insertion saturates at D = 15.9 (target 25 unreachable), so p7 runs D = 5/9/15 only. Real p7 D = 4.66.

L6 OUTCOME (02:50Z 11-16): see results/l6/READING.md. T2 refutation threshold fired (p199 gain 0.317). T1 law-form prediction failed (c 0.72->0.93 on random triples). Reading: fixed-c law refuted; no evidence of mixed structure; refuted-branch construction lane NOT opened (premise removed by T1).

L6 T1 REPLICATION (04:18Z): p19 D=9 random instances with generator seeds 2,3 (seed 1 gave 58 = real). Prediction: 57-59 each. If either is <= 55 or >= 61, the posted 'real = random at matched D' claim needs a correction.
L6 T1 REPLICATION OUTCOME (04:19Z): generator seeds 2 and 3 give 58, 58 and 59, 59 (LNS 60 s, two search seeds each). Inside the 57-59 band. Real 58. Claim stands; no correction needed.

=== L7 PRESTATED (05:58Z 11-16, before any L7 run) ===
Instance: radon U_{p,-1}, side n=2p. Hard lines = all pure lines + all lines with >=4 points. Soft lines = mixed 3-point lines.
Controls: same hard lines, soft lines replaced by random linear 2+1 triples at the real D (code/l6/t1gen.py logic), seeds 1, 2.
Quantity: v(s) = min over A, |A|=s, A hard-feasible, of #soft lines inside A; found by swap annealing (code/l7/sa.cpp), 3 seeds.
Report r = v/(D*n). Sizes s = 3p (1.5n) and round(3.6p) (1.8n); p = 43, 83, 131, 199.
Model expectation (pseudorandom soft lines): r rises toward (4/3)(beta/4)^3 = 0.070 (1.5n) and 0.122 (1.8n) as D grows.
Readings (from MR174 decisions):
- flat: r(p199) >= 0.5 r(p43) on real and both controls, at both sizes.
- falling: r(p199) < 0.25 r(p43) on the real carrier only (controls flat) -> construction lead.
- solver-limited: at p43 the 3 seeds differ by more than 10% of their mean, or the adequacy check fails.
Adequacy: at p43, s=121 (P99's no-three witness size) the annealer must reach v=0 on the real carrier in at least 1 of 3 seeds.
Prior (mechanism: L5/T1 found real = random at matched D): flat 70%, falling 5%, solver-limited 15%, other 10%.
=== L7 DEVIATIONS (06:11Z 11-16, after the first 45/75 s runs, before any further run) ===
Observed: (a) s=3.6p gave no hard-feasible state in 9/9 runs at p43 and 3/3 at p83. Hard-only LNS (pure + >=4-pt lines, 20 s)
max = 152, 288, 444, 671 at p43..199, all < 3.6p (results/l7/hardmax.txt). v(1.8n) is undefined as designed. Reading for the
1.8n arm: OTHER (not in the prestated list): no set of size 1.8n respecting only the hard lines was found at any p.
(b) p43 s=129 real seeds 5,8,7 (spread 0.45), ctrl1 7,6,5, ctrl2 3,3,3. The prestated SOLVER-LIMITED reading fires at 45 s.
Note: at v~6 one unit is 15%, so the 10% rule cannot pass unless seeds tie; I keep it as written for the 45 s run.
Diagnostic D1 (R4): p43 real s=129, 3 seeds x 300 s. Pass if the 3 minima are within 1 of each other AND min <= 45 s min.
If D1 passes: rerun the 1.5n arm only (p43..199, real + 2 ctrl, 3 seeds) at budget 300 s x (p/43), capped by the 2 CPU-h total;
readings flat/falling as prestated, at 1.5n only. If D1 fails: the curve is solver-limited; post the question without it.
Arm 1.8n replacement: none. I report the hard-only finding instead, labelled LNS lower bounds (not upper bounds).
=== L7 D2 PRESTATED (06:13Z 11-16, before running) ===
Question: does a certified integer upper bound below 1.8n (155 at p43) already exist at finite p? P99 MIP bounds are 1 s only.
Run: CP-SAT, 2 workers, 600 s each, on p43 g=-1 (a) hard-only graph tmp/l7_h43.txt, (b) full graph (hard + soft). Report best
objective and proven bound. Hard-only bound is also a bound on alpha(U), since hard-only is a relaxation.
Readings: bound < 155 on (a) or (b) -> finite certified alpha(U_43) < 1.8n (finite fact only, says nothing asymptotic).
bound >= 155 on both -> no finite evidence; the 1 s MIP picture stands at 10 min. Prior: (a) or (b) < 155: 35%.
deviation: 1 CP-SAT worker per run (satgraph.py default), not 2; 3 SA threads busy
=== L7 D1 RESULT + ALLOCATION (06:20Z 11-16, before any p131/p199 run) ===
D1: p43 real s=129, 300 s x 3: minima 6, 5, 5. Within 1, min 5 <= 45 s min 5. PASS. At p43, 45 s is at the 300 s level.
Spent: 4220 of 7200 CPU-s (SA 2940, hard LNS 80, D2 1200). The prestated 300 s x p/43 cannot fit the cap.
Allocation (cap binds): keep p43, p83 runs as done (45 s, 75 s, 3 seeds; p43 real also has the 300 s D1 runs).
New: p131 at 150 s, p199 at 220 s, real + 2 ctrl, 2 seeds each = 2220 CPU-s. Total 6440 of 7200.
Adequacy at large p is NOT established by D1. Extra reading: if the 2 seeds at p199 differ by > 10%, the p199 point is
solver-limited and I report the curve to p131 only. Readings flat/falling apply to r(p_max)/r(p43) at 1.5n.
=== L7 D2 RESULT (06:23Z): CP-SAT 1 worker 600 s, p43: hard-only val 150 bound 160; full val 115 bound 160. Both bounds >= 155: no finite certified alpha < 1.8n. Reading: 1 s MIP picture stands at 10 min (1 worker). ===
=== L7 H1 PRESTATED (06:27Z 11-16, before running) ===
Question: is the hard-only (pure + >=4-pt lines) LNS trend 1.767/1.735/1.695/1.686n (20 s, seed 1) a budget artifact?
Census: mixed >=4-pt lines are 2+2 quartets 307/657/1153/1848 of 325/690/1205/1926 (p43..199), per side 3.78..4.84.
Run: lns on tmp/l7_h199.txt, 150 s, seeds 2 and 3 (sequential, 1 thread). Then p83 hard-only 75 s seed 2.
Reading: at p199, if max(150 s) >= 1.8n (717) -> hard-only is NOT below 1.8n; drop the reframing.
If 150 s gains > 0.02n (8) over 671 -> trend solver-limited, report as such. Else -> hard-only LNS < 1.8n stands (LOWER
bounds only; no upper bound claimed). Prior: stands 70%, gain >0.02n 25%, >=1.8n 5%. CPU: 375 s (L7 total 6815 of 7200).
=== L7 H1 RESULT + Q1 PRESTATED (06:40Z 11-16) ===
H1: p199 hard-only LNS 150 s seeds 2,3 = 677, 678 (gain 7 < 8 = 0.02n over 671); p83 75 s = 288 (= 20 s). Reading: hard-only
LNS < 1.8n STANDS (lower bounds 1.767/1.735/1.695/1.704n at p43..199). 1.5n readings: FLAT fires (r rises 4.4x real, 4.8x/8.6x ctrl).
Q1: keep pure lines and mixed >=5-pt lines; replace the 2+2 quartets by the same number of random linear 2+2 quartets (no pair
reused), seed 1. LNS 20 s seed 1 at p43..199 (80 CPU-s). Compare with real 152/288/444/671 (20 s seed 1).
Readings: |ctrl-real| <= 2% of real at all 4 p -> quartets random-like. real > ctrl by > 3% at >= 3 p -> real quartets
easier to satisfy (structure). real < ctrl by > 3% at >= 3 p -> real harder. Else mixed. Prior: random-like 55%, easier 25%, harder 10%.
=== L7 Q1 RESULT + Q2 PRESTATED (06:47Z 11-16) ===
Q1 (uniform random linear 2+2 quartets, LNS 20 s seed 1): ctrl 155/295/465/696 vs real 152/288/444/671 at p43..199; real lower
by 1.97/2.43/4.73/3.73%. Prestated reading: MIXED (>3% at only 2 of 4 p). Both selections at p199 re-verified, 0 violations.
Caveat: Q1 ignores per-point quartet degrees. Q2: rewire_linear (code/l5/rewire.py) on the 2+2 quartets with keep = other hard
lines: per-point degrees exact, linear. Seed 1, LNS 20 s seed 1, p43..199 (80 CPU-s; L7 total 6975 of 7200).
Readings: real < Q2 by > 3% at >= 3 p -> real quartets harder than degree-matched random (structure; lead = which quartets
bind). |Q2-real| <= 2% at all p -> Q1 gap was a degree effect. Else mixed. Prior: degree effect 45%, harder 30%, mixed 25%.
=== L7 Q2 RESULT (06:49Z): degree-matched quartets (rewire_linear, 0 bad) LNS 153/291/448/675 vs real 152/288/444/671: within 2% at all 4 p (+0.66/+1.04/+0.90/+0.60%). Reading: Q1 gap was a DEGREE EFFECT; real quartets random-like given degrees. L7 CPU total 6975 of 7200: compute stopped. ===
=== L7 E1 PRESTATED (07:02Z 11-16, before running; new item, own budget 15 CPU-min, outside the spent L7 2 CPU-h) ===
Question: does the hard-only relaxation (pure + >=4-pt lines) of U_{p,-1} show an integrality gap at small p, exactly?
Run: p = 7, 11, 19, 23 (all 3 mod 4). LP_h = max sum x, 0<=x<=1, sum_L x <= 2 on hard lines (GLOP). alpha_h = CP-SAT 1 worker,
120 s per p; report OPTIMAL or bound. Readings: alpha_h <= LP_h - 0.05n at every p solved OPTIMAL -> finite gap present.
LP_h - alpha_h < 0.02n at every OPTIMAL p -> no finite gap at small p (the gap, if any, is asymptotic). Else mixed.
Fewer than 2 OPTIMAL -> solver-limited. Prior: gap present 60%, no gap 20%, mixed 10%, solver-limited 10%.
=== L7 E1 RESULT (07:11Z): exact alpha_h (OPTIMAL) 26/41/84 at p7/11/23 = 1.857/1.864/1.826n; LP_h 1.929/1.909/1.886n; gap 0.071/0.045/0.060n. p19 FEASIBLE 67, bound 71. Reading: MIXED (p11 gap 0.045n < 0.05n). Note alpha_h(p23) > 1.8n exactly: the post's 'integer gap already present' rests on LNS lower bounds at p>=43. Adequacy check A2 (prestated now): LNS 20 s seed 1 must reach 41 (p11) and 84 (p23); else LNS hard-only values are unreliable. ===
=== L7 A2 RESULT + E1b PRESTATED (07:12Z): A2 PASS (LNS 20 s = exact 41 at p11, 84 at p23). E1b: p31 CP-SAT 150 s + hard-only LNS 20 s (E1 total ~690 of 900 CPU-s). Reading: alpha_h(31) or LNS >= 112 (1.806n) -> crossing below 1.8n lies in (31,43); < 112 -> in (23,31]. ===
=== L7 E1b RESULT (07:21Z): p31 CP-SAT 150 s FEASIBLE 108, bound 115 (1.855n); LNS 20 s 110 = 1.774n < 112. Reading: crossing below 1.8n (by LNS) lies in (23,31]. p19 LNS 67 = CP-SAT 67 (bound 71): 1.763n, non-monotone vs p23 exact 1.826n. E1 CPU ~710 s of 900. ===
=== L7 C1 PRESTATED (07:40Z 11-16, before running; budget 15 CPU-min) ===
Hard-only instance on U_{p,-1}. For each point v: S_v = union of hard lines through v. a_v = exact alpha of the instance induced
on S_v (constraints L cap S_v, |L cap S_v| >= 3), CP-SAT. Cut: x(S_v) <= a_v. LP_c = LP_h + all cuts (GLOP).
closure = (LP_h - LP_c)/(LP_h - alpha_h), at p = 7, 11, 23 (exact alpha_h); at p43 report LP_c/n only (alpha_h unknown).
Readings: closure >= 0.5 at all 3 p -> local cuts carry the gap (discharging route plausible). closure <= 0.15 at all 3 p ->
gap is global. Else mixed. Also report whether LP_c(p43)/n < 1.8. Prior: global 45%, mixed 35%, local 20%.
=== L7 C1 RESULT + C1b PRESTATED (07:41Z 11-16) ===
C1: closure 0.000 at p7, 11, 23 (LP_c = LP_h; 56/88/184 cuts all OPTIMAL). Prestated reading GLOBAL fires, but the test is weak by
construction: lines through v meet only at v, so alpha(S_v) ~ 2k, which the LP already implies. Not evidence about geometry.
C1b: pair cuts. For every pair u,w on a common hard line, S_uw = union of hard lines through u or w (lines through u cross lines
through w inside S_uw, so internal conflicts exist). Cut x(S_uw) <= alpha(S_uw). Same closure readings as C1, p = 7, 11, 23.
Prior: global 45%, mixed 35%, local 20%. Budget: 10 CPU-min.
=== L7 C1b RESULT (07:42Z): pair cuts (400/683/1423, all OPTIMAL) closure 0.000 at p7, 11, 23. Reading GLOBAL. Diagnosis before a third family: absolute gaps are 1, 1, 2.76 points; measure LP_h vs LNS alpha_h at p43..199 (GLOP, seconds) and run the prestated C1 star cuts at p43. ===
=== L7 C1 p43 + LP_h (07:43Z): p43 star cuts (344, all OPTIMAL) LP_c = LP_h = 1.8681n, closure 0. LP_h/n 1.862/1.868/1.862/1.853/1.853 at p31/43/83/131/199; LP_h - LNS_h = 0.088/0.101/0.127/0.158/0.149n (upper bound on the gap). Local cut test closed; no third family. ===
=== L7 C2 PRESTATED (08:09Z 11-16, before running; prompted by lattice S14rZRCqnKvnWyxp; budget 20 CPU-min) ===
C1 tested only cardinality cuts. C2 tests ALL valid inequalities on each support S: separate x*|S from conv(feasible subsets of S)
by column generation (master GLOP min shortfall; pricing exact CP-SAT max W.F with integer W). Any W with W.x* > max_F W.F is a
valid violated cut. Outer loop: re-solve LP_h + cuts, repeat until no support separates (round cap). Supports: stars S_v.
p = 7 (LP_h 27, alpha_h 26) and p = 11 (LP_h 42, alpha_h 41); pairs at p7 if time allows.
Readings: final LP <= alpha_h + 0.5 -> star-local hull closes the gap (local route plausible). final LP >= LP_h - 0.01 -> star-local
hull null (every star restriction of some LP optimum is integral-hull feasible). Between -> partial. Prior: null 55%, partial 30%, closes 15%.
=== L7 C2 RESULT (08:23Z 11-16) === code/l7/hull.py; results/l7/c2_p7.txt, c2_p11.txt, c2_p7pair.txt
Stars, converged (a round with 0 new cuts, every star master pi = 1.0, iteration cap never hit):
  p7: 21 cuts over 6 rounds, LP 27.0 at every round (LP_h 27, alpha_h 26). p11: 12 cuts over 6 rounds, LP 42.0 at every round (LP_h 42, alpha_h 41).
  Reading: NULL fires. The final LP optimum x* has every star restriction inside conv(feasible subsets of the star), so the
  star-local integer-hull relaxation has value exactly LP_h at p7 and p11. This covers all valid inequalities on the star supports, weighted ones included.
  The cuts that were found only moved the LP between alternative optima.
Pairs at p7: timeout 420 s after 3 rounds, 348 cuts, LP 27.0 at every round; NOT converged, so no reading (it is consistent with null).
Deviation: the pair arm was capped by wall time (one process, about 7 CPU-min). Total C2 CPU is about 8 min, within the 20-min budget.
=== L7 C2 CERTIFICATE NOTE (08:27Z) ===
Why min_mu = 1.0 certifies the null. The master LP over the generated columns lower-bounds the full separation value
min_{w>=0, w.x*=1} max_F w.F, because it has fewer columns. So master pi >= 1 means x*|S lies in conv(feasible F of S), with no
rounding risk in that direction. Rounding W to integers affects only which cuts are found, not this certificate. Dropping the
points with x*=0 from S is exact, since the hull is down-monotone. Remaining caveat: GLOP tolerance (pi is printed to 5 decimals).
=== L7 C2b PRESTATED (08:36Z 11-16, before running; fills a limitation of write-up section 6; cap 10 CPU-min, one process) ===
Same hull.py, stars, p = 23 (LP_h 86.764, alpha_h 84 OPTIMAL), rounds cap 15, timeout 600 s.
Readings use the C2 thresholds: final LP <= 84.5 closes; final LP >= 86.754 null; else partial. If it times out before a round with 0 new cuts: no reading.
Prior: null 70%, partial 25%, closes 5%.
=== L7 C2b RESULT (08:38Z) === results/l7/c2_p23.txt
Converged in 2 rounds and 26 s. Round 0: 2 weighted star cuts. Round 1: 0 new cuts, every star master pi = 1.0.
LP 86.7641 -> 86.7593 (drop 0.0048, i.e. 0.0001n; alpha_h = 84). Reading: NULL fires (86.7593 >= 86.754 threshold).
Unlike p7 and p11, the star hull is strictly tighter than LP_h here, by 0.2% of the gap (0.0048 of 2.764).
=== R1 PRESTATED (09:01Z 11-16, before running; MR175 D0; budget 1.5 CPU-h, at most 5 threads) ===
Full problem at p199, all lines (cap 2). Real baseline from L6: LNS 900 s, seeds 1-3 = 525, 519, 516 (results/l6).
Controls: code/l7/mk.py ctrl seeds 1 and 2 (hard lines kept, mixed 3-point lines replaced by random linear 2+1 triples at real D).
LNS 900 s, seeds 1-3 on each control, same binary code/l6/lns. Every selection verified against its own graph.
Readings: STRUCTURE if the real best (525) exceeds every control seed AND real mean - control mean >= 8 (2% of side 398).
RANDOM if the control best >= the real best - 2 (i.e. >= 523). Otherwise INCONCLUSIVE.
Prior: random 60%, inconclusive 30%, structure 10%.
Caveat known before running: L6's real runs used the L6 graph (code/l6/mkgraph.py), not the l7 format. Lines are the same object; I check that the real l7 graph, converted, gives the same line count.
=== R1 RESULT (09:32Z 11-16) === results/r1/sel_c{1,2}_s{1,2,3}.txt, all 6 verified against their own graphs (0 violations)
Controls at 900 s: c1 = 518, 515, 511; c2 = 517, 511, 515. Real (L6) = 525, 519, 516.
Real best 525 > every control (max 518): yes. Mean difference 5.5 points (1.4% of side 398) < 8: no. So STRUCTURE does not fire.
Control best 518 < 523: RANDOM does not fire. Reading: INCONCLUSIVE.
Descriptive: one-sided permutation test on means, 4/84 = 0.048. That is a small real advantage (about 1.4% of the side) at D = 26.9,
in the same direction as the 1.5n annealer gap at p199. Goal relevance: none as a construction (1.32n). Per MR175 D0 the
inconclusive branch only scopes paper Section 3; no structure lane is opened.
=== C2 EXACT CERTIFICATE (09:49Z 11-16, prompted by lattice 8yzmCejx17q8ZQn_) ===
Rerun hull.py at p7, p11, p23 (now dumping the final x, cuts and last-round columns). Then cert.py checks in exact rationals:
the rationalised x satisfies every hard line, and for every star it is dominated by a convex combination of verified-feasible
columns. A pass at all stars certifies alpha_star >= sum(x) exactly. Failure modes: rationalisation breaks feasibility, or the
lambda rounding is inexact. Then I report the certificate as numerical only. The p199 permutation p-value is WITHDRAWN (nested design, lattice).

[10:40Z 11-16] D-run dual: rational dual upper bound on LP_h at p=7,11,23 (code/l7/dual.py), per lattice q-bfRROxHvFYEAyx item 3.
Dual: min 2*sum y_l + sum z_i, y,z >= 0, sum_{l ni i} y_l + z_i >= 1. Rationalise GLOP y, set z_i = max(0, 1 - sum y) exactly.
Prediction: certified UB within 1e-6 of printed LP_h (27.0, 42.0, 86.76 approx). Reading: if UB - alpha_star_LB <= 0.005 at all three, the
"at most 0.005" claim becomes rigorous; else state the decreases numerically only. Decision it changes: abstract/s6 wording.
[10:43Z 11-16] READING dual: LP_h <= 27, 42, 86.76408377445 exact (UB - GLOP <= 2.2e-10). LP_h - alpha_star <= 0, 0, 0.0049004 (< 0.00491).
Prediction met. "at most 0.005" now rigorous on both sides; manuscript s1/s6/s8 and abstract revised per lattice q-bfRROxHvFYEAyx.
