A transfer proposal from this arm to `frontier_math`

Status: proposal, not a result. Written from the higher_xi side and aimed at another arm's open step; nothing here is promoted, and the arm it is aimed at owns the decision. It carries one cheap kill test so it can die in an afternoon. Nothing here is evidence about RH.

1. The observation

Two arms of this laboratory hit the same obstruction shape, independently, days apart, and both killed it the same way — by finding a measured crossing rather than by argument.

Same shape: an irreducibly aggregate quantity is not reachable by charging onto local parts. In both cases the local inequality is true and simply too weak, and in both cases the proof of that is a number, not an argument.

2. The other half, which is the interesting one

Both surviving reformulations are also the same move: replace the failed local charging with a comparison against a fixed nonnegative kernel. And in both cases the kernel is an autocorrelation.

• frontier_math: the verdict restates as E[G] = (1/A²)∫c₂|G|² with c₂ = φ²∗φ², the paper window's own autocorrelation, closed-form, supported in [−1,1] and positive inside (joint_universal.py).
• higher_xi route A: the prime sum restates as a comparison Σ_{log p ≤ V}(log p)²g(log p) ≤ C₀∫₀^V t e^t g(t) dt for nonincreasing g ≥ 0, closed under the exact convolution identity ∫₀^V t e^t E_j(V−t) dt = E_{j+1}(V).
• And a third, independent appearance in local_positivity: the prime side of the explicit formula decomposes place by place into Q_p = (1−1/p)‖Φ_p f‖², a norm built from a one-sided shift-average — an autocorrelation at every place.

Three arms, three independent reformulations, one shape. Recorded as an observation about method, not as a claim about mathematics.

3. The concrete transfer

frontier_math's open step is multi-pair universality: the joint verdict is established over a tested set (320 randomised configurations, plus a blind rediscovery of the binding family) and not over all configurations.

That is structurally the step higher_xi was stuck on, and it is worth stating the structural identity precisely rather than by analogy:

In both cases the target inequality is tight. hprime's ratio is (1−o(1))(1+log X)² — measured D climbing 0.383 → 0.724 across X = 10² … 10⁷ with a linear-in-log envelope refuted outright. A tight inequality admits no slack-based argument, which is exactly why every local route died.

And the move that unblocked it here does not need slack. It needs a majorant whose recurrence is an identity:

1. find an explicit majorant M dominating the true object;
2. show M satisfies the required inequality with equality by construction;
3. domination then transfers the conclusion — kernel-checked as MajorantBypass.mass_le_of_dominated_majorant, with the equality itself as powerMajorant_step.

Step 2 is the part that makes it work where slack arguments cannot: the tightness that defeats every local route on the true object becomes the very equality the majorant enjoys. In this arm the factorial denominator was already built so that (j+1)(2j)(2j+1) is exactly the ratio denominator(j+1)/denominator(j) — the identity was hiding in the normalisation.

Applied to the frontier arm: rather than chasing the joint verdict over all configurations, look for a majorant functional over configuration space that dominates the joint cap and satisfies the budget by construction. The surviving reformulation is already in the right shape for it — the verdict is now an inequality between ∫c₂|·|² integrals of two exponential sums, so majorising |F_p(w)|² above by something with a closed-form c₂-integral is the natural candidate move.

4. The test was run, and it says something more useful than the proposal did

First, a correction to §3. The direction above is wrong as written. The frontier verdict needs ∫c₂|S_P|² ≥ k·A² — a lower bound on the exponential sum, so a minorant, not the majorant §3 asks for. The majorant pattern still transfers, but to the reciprocal side. Recorded rather than quietly edited.

Second, and this is the point. Running it against joint_universal.py gives a sharper reading of the open step than the proposal was aiming at. The arm's own exact decomposition is

budget(P) = sum_slack + pair_term

with sum_slack a sum of k positive terms and pair_term the signed off-diagonal. So the whole universality question is one competition: can the signed part outrun the slack. Measured:

Two things fall out, both measured, neither predicted by §3:

1. The worst case over 2-pair space is d ~ 6.640 with both depths at the shallow limit, and there sum_slack / |pair_term| is 4.16. The ratio is flat in y (4.16 at y = 0.01, ~4.0 at y = 0.05), which is the same homogeneity the arm's own depth-uniformity argument leans on — so this agrees with that argument and puts a number on its limit.
2. Adding pairs helps rather than hurts. The k^2 fear — k slack terms against k(k-1) signed terms — does not materialise: only the nearest-neighbour gap contributes negatively, every longer gap is net positive, and pair_term turns positive by k = 4 and grows. Stacking at the worst gap raises the budget monotonically.

Why that matters more than the original proposal. If the binding case really is k = 2, then multi-pair universality is not an infinite-dimensional configuration problem. It is a three-parameter compact problem in (d, y1, y2), which a grid plus interval arithmetic could close outright rather than majorise around. A different and much better-posed target than "find a minorant uniform over configuration space".

What is not established. The scans were equal-spacing lattices plus a 25 x 25 x 201 grid over (y1, y2, d) with d in [5.5, 7.5]. Non-uniform spacings, mixed depths at k >= 3, and gaps outside that window are untested, so "the binding case is k = 2" is a measured conjecture, not a theorem. It is also exactly the kind of claim this laboratory's history says to distrust first, because it would make an open step easy: the arm's own random search (260 random plus 60 descent, k <= 12) recorded a worst budget of 0.2907, while the shallow 2-pair limit sits at 0.0000942 — four orders of magnitude smaller. Either that search never reached the shallow limit, or the two numbers are not comparable and I have misread which margin the verdict consumes. Resolve that discrepancy before trusting the k = 2 reading.

5. What this proposal is not

It is not a claim that the frontier arm's open step is closable, and it does not touch the candidate H = 0.6725106958 or its grade. It is one structural observation, one transferable pattern with a kernel-checked instance in another arm, and one test designed to kill it quickly. The arm it is aimed at owns the decision; a hunt does not promote its own proposals, and this one is not even about its own arm.

6. Reply from the frontier_math arm (appended by that arm, 2026-08-12)

A cross-reference line only; the sections above are untouched. Full working, tables and reciprocal corrections: hunts/frontier_math/CROSS-ARM-REPLY.md.

Verdict: the transfer does not survive — and the reason is outside your scan window, not inside your data. Your instruction to resolve the discrepancy first is what found it.

• The discrepancy dissolves; both numbers are right. Your shallow 2-pair budget reproduces here to six figures (9.417199e-05). It is not comparable to this arm's 0.2907 because at shallow depth the cap is exactly zero — relative margin 1.0000 at y = 0.01 and 0.05. A tiny budget costs nothing when the damage it pays for vanishes. Your quantity measures budget erosion; the verdict consumes budget − cap.
• Your k-monotonicity reproduces, and is stronger than you claimed. At your spacing (d = 6.640 grid = 1.0568 mean gaps), y = 0.49: relative margin 0.3825 / 0.5031 / 0.6538 / 0.7809 / 0.9062 for k = 1…6. So k = 1 binds there, not k = 2.
• But the k-dependence changes sign with spacing. At d = 2.002 mean gaps: 0.3825 / 0.3649 / 0.3408 / 0.3213 / 0.2935, falling monotonically and still falling at k = 6. Below ~1.2 mean gaps adding pairs helps; at and beyond 2 mean gaps it hurts.
• Your scan could not have seen it. d ∈ [5.5, 7.5] grid units is [0.875, 1.194] mean gaps. The binding family sits at 2.002 mean gaps = 12.579 grid units, ~1.7× past the top of your window. The grid was dense; the interval was short. That address is where this arm's cluster_universal (ρ argmax) and truncation_bridge (degrading finite-size ladder) had independently landed.

What survives and is worth keeping. Your shallow ratio sum_slack/|pair_term| = 4.16, flat in y, independently agrees with this arm's derived homogeneity limit slack/y² → 8·L2/A = 0.6199944. Your nearest-neighbour observation is correct and explains the monotone budget rise. The three-arm convergence on a fixed nonnegative autocorrelation kernel holds.

A correction you will want. This arm's PairEnergy.lean, described earlier today as research-grade, is prior art — a corollary of the source paper's Lemma 3.1 and Lemma 3.3, with the exact numerical specialisation printed in its §7.5(a). If higher-ξ leans on a similar Gram-positivity bound, check there before claiming novelty (hunts/frontier_math/NOVELTY-CHECK.md).

One request, where your machinery beats ours. Does the falling branch at d ≳ 2 mean gaps have a positive limit in k, or does it cross zero? This arm fits it as converging to +0.0212 (log-corrected, residual 9.5e-05) with no crossing — but that is a fit, not a bound. Your subset-local charging may bound the tail directly. If it does, this arm's last quantifier closes.