Plan file: rung3_plan2.json. The naive boxed-s lower-bound evaluation is gone; the small-square frontier is now certified by point enclosures on a grid + a Lipschitz bound L, where L = M/(w2-w) comes from boxed-s upper bounds on a big square (width tolerated, not fatal). All coordinates are exact rationals (denominators divide 2^12 * 10^6); every certificate-side inequality was re-checked in exact Fraction arithmetic by an independent pass (v2_verify.py, reads only the JSON): 3137 checks, 0 failed.
Headline: small square w = 3/64 around (808517/1000000, 85699348/1000000), eps' = 1/2000; big square w2 = 7/32; L = 16 (= "16/1"), M = 25499/10000; 100 unique grid points (104 entries; the 4 corners are shared pairwise, duplicate entries carry identical data), 110 big boxes, centre K = 361. Total certified-term cost = 77675 (grid 49360 + big 26510 + centre 1805) -> wall-clock ~ 18.1 h at 0.84 s/term on 4 cores.
Chosen parameters and why
Scanned w in {1/32..1/8}, w2 in {3/16..5/16}, M-target in [0.6, 4.0], eps' in {3e-4, 5e-4, 1e-3} with greedy covers under the stated width model. The surface is flat near the optimum (76.6k..80k over a wide region):
- w = 3/64: the grid-point count is ~invariant in w (N ~ 5.5 L because both the frontier floor and the allowed gap scale with w), but per-point K falls with w while L = M/(w2-w) rises as w eats into the Cauchy gap; w = 3/64 balances this against w2 = 7/32 (gap 11/64). cre - w = 0.76195... > 1/2 (exact check). Frontier floor max(|Re|,|Im|) = 0.0492; measured max |DH'| on the small frontier = 1.30.
- w2 = 7/32: pushing w2 left lowers sigma on the big square's left edge (re_lo = 0.5898), and both Sigma(K) and the tail exponent degrade fast below sigma ~ 0.6 — at w2 = 5/16 the big cover costs 4-10x more. Pulling w2 in shrinks the Cauchy gap. 7/32 (with sigma_lo down to 4/7 on the left edge) is the measured sweet spot.
- M-target = 2.5 (=> M = 25499/10000, L = 16): the explicit trade grid ~ c*L vs big-cover ~ falling in M; the scan minimum is at 2.5 with 76,605 planned terms (2.2 -> 78.9k, 3.2 -> 79.8k). Exact build lands at 77675.
- eps' = 1/2000: 768x above ||DH(centre)|| = 6.51e-7; the centre now keeps a >= 25% margin (K_c = 361, r_c = 3.736e-4, margin 1.257e-04 = 25.2% of budget). eps' = 1e-3 saves ~400 centre terms but costs ~ the same on the grid: flat.
The plan
- Grid (26 points per segment, incl. endpoints; corners shared): bottom K in 85..113, top 85..107, left 101..114, right 85..94. Adaptive gaps h_i (larger where ||DH|| is larger): h from 141/32768 up to ~3x that. Per point: sigma_lo = best k/d (d<=8) <= Re(p) — 3/4 on most of the frontier, 4/5 where Re >= 0.8; point tail r_i ~ (maxcomp - eps')/(sigma_lo+3), K_i minimal for it. Cell requirement (exact): min(beta_i, beta_i+1) - L h_i/2 - eps' >= (1/4)(min beta - eps'), with beta_i backed by mpmath dps-40 point values (pred_beta) minus a 2x sampling slack (2e-11).
- Big boxes: bottom 24, top 23, left 56, right 7 (the low-sigma left edge needs the most; the right edge, sigma_lo = 1, needs 7). K in 17..61. Per box predM = max|Re| + max|Im| + 2*width + 2*r (see model note below), all predM <= M, and L*(w2-w) >= 1.1 * max predM (exact; headroom lands at 1.100).
- Centre: K = 361, sigma_lo = 4/5 (a/b = 14/5), r = 205390733/549755813888; ||DH(c)|| + slack + r < eps' with 25.2% margin.
Verification (3137 checks, 0 failures)
- Structure (exact): grid covers each segment end-to-end, gap_next matches successor differences exactly, last gap null; big boxes tile each big segment exactly with shared endpoints; degenerate coordinates exact; all denominators divide 2^12 * 10^6; corner duplicates agree field-by-field; every K <= 450; re_lo > 0 on every big box; im > 0 everywhere.
- kappa_hi >= kappa proved exactly (integer-sqrt enclosures at 10^-25).
- Exact chains everywhere (grid, big, centre): sigma_lo <= re, denominator <= 8; a/b = sigma_lo + 2 reduced; P^b (5K+1)^a >= 1 in integers; NS/NS1/NS2 >= normBound exactly; r >= (5/8)(3+kappa_hi) NS NS1 NS2 P b/a.
- Grid betas: fresh mpmath dps-40 evaluation at all 100 unique points; pred_beta reproduced to < 1e-9; cells re-checked in exact rationals with beta reduced by 2x sampling slack. Worst three cell margins: 0.270 (g_left_22->g_left_23), 0.270 (g_top_07->g_top_08), 0.270 (g_top_23->g_top_24) — all >= the required 0.25 of budget (and >> 0.25 eps').
- Big boxes: fresh 3x + midpoint resample per box (max|Re|, max|Im| with a 2x derivative-based sampling guard), Sigma re-summed independently (math.fsum); every fresh predM <= stated predM <= M; the numpy EM sampler was spot-checked against mpmath dps-40 at 3 points per box (330 points, worst |diff| = 4.1e-14) — this matters because the big frontier reaches sigma ~ 0.59 where the sampler had not been validated before.
- Centre at dps 50; margin 25.2% of budget (>= 25% required).
Width-model plausibility (mandated check)
Model: enclosure width over a 1D box of extent delta ~ 1.30 * delta * Sigma(K) + 5K 2^-50 + 1e-9, Sigma(K) = sum_{m<5K, m%5!=0} m^-sigma ln m. Derivation: the kernel's series is sum a_m m^-s (|a_m| <= 1); over the box, each term sweeps an arc of length <= delta * ln m * m^-sigma_min (mean value in s); summing gives delta * Sigma. Numbers at two boxes / two K each:
- worst box B_top_11 (sigma_lo = 2/3, delta = 0.01410): K = 44: Sigma = 41.9, width = 0.768; K = 88: Sigma = 63.6, width = 1.165. Direct chord-sum cross-check at K = 44: sum |m^-s_a - m^-s_b| = 0.500 vs delta*Sigma = 0.591 (ratio 0.85) — Sigma is near-sharp; the 1.30 factor is the real safety.
- cheap box B_right_06 (sigma_lo = 1, delta = 0.02179): K = 17: Sigma = 7.9, width = 0.224; K = 34: Sigma = 10.6, width = 0.299; chord ratio 0.92.
The width term dominates predM exactly as anticipated: worst box breakdown predM = X + 2*width + 2*r = 0.384 + 1.536 + 0.580 = 2.4999 (X = true max|Re| + max|Im| incl. sampling guard).
Hostile-referee notes
- The normBound reading of predM is the load-bearing modelling choice. The coordinator's literal model ("|DH| true max + halfwidth + r") gives max predM = 1.39 -> L = 8.87. I used the L1-conservative reading (normBound sums per-component maxima, each padded by the full width and by r): max predM = 2.50 -> L = 16. Ratio 1.8x. I declared the conservative L because an underdeclared L fails the kernel check L(w2-w) >= M outright. If a pilot run of one big box shows the kernel's normBound tracks the literal model, re-planning with L ~ 9 cuts the grid to ~30k and the total to ~56k terms (~13 h). The failure mode of my choice is cost, not soundness.
- The width model assumes per-term arc/mean-value enclosures. If the kernel encloses m^-s by naive interval exp/log in re/im components, its width can exceed delta*Sigma (extra wrapping), M could exceed 25499/10000, and the L check fails — loudly, not silently. The 1.30 factor plus the 1.02 (M) and 1.10 (L) headrooms absorb ~43% combined; the chord check says Sigma itself is only ~10-18% conservative, so the total cushion over a mean-value kernel is ~1.5-1.7x, but a genuinely naive kernel would blow through it. Verify the evaluator's form before spending 18 hours.
- The big-square detour buys less than it appears. Boxed-M total = 77675; the closed-form fallback L = 40 (no big square at all) costs 88,875 at its own best w (1/8), i.e. only ~14% more, with ~110 fewer kernel lemmas and no M/L machinery. If proof-engineering time matters more than 2.4 h of compute, the fallback is defensible. Conversely "more/smaller big boxes" does NOT fix the width model: pushing M-target to 1.0 costs 242k terms (608 boxes) — cost/length diverges as delta -> 0. The scan table in
v2_buildoutput records this. - L = 16 vs measured |DH'| <= 1.31 on the small frontier: the architecture pays ~12x over the true Lipschitz constant (Cauchy over a thin annulus plus L1 norms plus width padding). Any future certified derivative bound |DH'| <= ~2 on the small square would collapse the grid to ~15 points and the total to ~10k terms. Worth a rung of its own.
- Sampling exposure is now structural, not incidental: cells need no interior samples at all (mean value covers between points; betas are mpmath-backed at the points). The only sampled maxima left are the big boxes' X terms, guarded by 2x the derivative-based gap bound (9.7e-05) and sitting under the 1.30/1.02/1.10 stack. The numpy sampler is validated to 4e-14 against mpmath dps-40 at sigma down to 0.59.
- Deviation from the schema note:
centrealso carries "sigma_lo"/"a"/"b" (= 4/5, 14, 5) so the kernel need not hardcode the centre's exponent pair; unknown-key-tolerant parsers are unaffected. - Wall-clock 18.1 h is the honest price under my conservative model; the realistic range after a pilot box is ~13-18 h, or ~2.4 h more for the L = 40 fallback with a much simpler proof.
Artifacts
rung3_plan2.json— the plan (kernel input).v2lib.py,v2_optimize.py— models, Sigma tables, (w, w2, M, eps) scan.v2_build.py/v2_build_extras.json— refinement + exact build.v2_verify.py/v2_verify_results.json— independent verification (3137 passed / 0 failed).