---

Record and current scope

This is the substantive review of the exact reviewed subject. First-person passages retain the original independent reviewer's voice. The review was made on 15 September 2026 by an independent reviewer in a fresh context distinct from the argument's author. The subsequent contract grade is by the commissioning grader, distinct from both author and reviewer. It addresses the original report contract and independence, not a second blind mathematical review.

The present filing was prepared by the exposed author. It does not provide independent correspondence review of this native rendering or a new verdict. The original report remains in working storage. The current proof text and research owner distinguish this retained assessment from any later mathematical revision.

Mathematical verdict

refutation-failed. The complete frozen argument survives this review of its unconditional existence statement, all its essential deductions, and its stated density consequences. I found no counterexample, omitted essential case, or unsupported bridge to the claimed bound.

This verdict relies on the exact standard external theorem interfaces identified below. It does not claim independent proofs of those external theorems, literature priority, a solution of E0158, or acceptance of any native claim at a numerical tier. A distinct report-contract grader and ordinary-clone integration remain necessary for any subsequent tier-bearing use. The original review did not itself complete corpus integration.

Subject, independence, and actual reading

Reviewer: the independent reviewer, in the fresh context identified above. I did not construct the subject or previously build on it. My mathematical context for this review comprised the admitted frozen subject, the expressly permitted L4 comparison/standing index, and the four permitted public source artifacts. No sibling verdict, private plan, author or collaborator advocacy file, other attack file, or excluded local proof was opened.

The frozen subject is the complete retained original, 667 lines. Its scope includes the standing, comparison, consequence and external-input sections. I read it completely; two observed hashes matched admission.

The frozen subject wiki/research/leads/balanced_auxiliary_prime_density/evidence/assets/loss_refinement_reviewed.txt as it stood on 2026-09-15T05:40:27Z carries the argument's own standing text at lines 21-26 and 661-667 (author-recorded, without independent review or acceptance; L4 and L2 at tier 0); a separately spawned materiality grader (model Claude Fable 5.1) ruled that exposure immaterial by the content test, because the text states only the null standing, implies no verdict, and the review's reasoning rests on its own rederivations and attacks rather than on that text.

The only other local mathematical page read was the complete L4 comparison index as it stood on 2026-09-15T02:18:25Z. I used it solely to compare the exact loss, conventions, scope, and recorded tier-0 standing. Its descriptions of author checks supply no mathematical support in this verdict. Its proof, L2, and all sibling claims and reports remained unopened. The current link is navigation; the date identifies the historical comparison actually read.

The separate operating read set comprised the complete admission and applicable organization, workspace, repository, anatomy, research, evidence, verification, approach-vetting and PDF-reading instructions. Those were operating context, not mathematical evidence. No additional mathematical read was authorized by their links.

I visually inspected nine selected full PDF pages. The initial direct web opens also exposed extracted text from the Erdős article and an introductory portion of Conrad's note; those automatic excerpts do not enlarge the proof-coverage claims below. No mathematical executable, solver, test, Lean command, package installation, or author run was used. The review did not modify the source proof or the native corpus and did not delegate mathematical work. Exact external-source identities and the substantive accounting limitations are retained below.

Full quantified restatement

Let $\gamma=\sqrt6-2$. The proposition asserts the existence of a single infinite set $A\subseteq\mathbb Z_{>0}$, a real $C>0$, and a real $x_0>e$, such that both statements hold:

1. For every integer $n$, there are at most two pairs $(a,b)\in A^2$ satisfying $a\le b$ and $a+b=n$. These are unordered representations; $a+a$ counts once.
2. For every real $x\ge x_0$,

$$|A\cap[1,x]| \ge x^\gamma\exp\{-C\sqrt{\ln x\,\ln\ln x}\}.$$

The set and constants are chosen before the real cutoff. The same set therefore has the following additional property: for each real $\alpha<\gamma$, there is a threshold $x_\alpha$ such that $|A\cap[1,x]|\ge x^\alpha$ for every real $x\ge x_\alpha$. In particular its counting-function logarithm has limit inferior divided by $\ln x$ at least $\gamma$. The latter observation does not assert a matching upper bound or a positive constant times $x^\gamma$.

The comparison is with L4's stated loss $\sqrt{\ln x}\,\ln\ln x$, with the same representation and cutoff conventions. The new loss exponent divided by that one is $1/\sqrt{\ln\ln x}$, which tends to zero. This is an improvement in the order of the explicit loss, without a larger power exponent or a public-record claim.

Setup and transfer checks

The elementary scale estimates are sufficient at their exact strength. The final half of the summands in $S_k=\sum_{j\le k}\lceil j\log_2(j+1)\rceil$ gives $S_k\gg k^2\log(k+1)$; the maximum summand gives the reverse bound. Thus $w_k=O(R_k)$, $R_k=o(S_k)$, and $R_k^2=O(S_k\log(S_k+2))$, where $R_k=k\log_2(k+1)$. Also $R_{k+1}=O(R_k)$. The strict increase of $w_j$ makes the half-open prime intervals disjoint. Each is nonempty, and every auxiliary-prime outcome satisfies the product bounds (3).

From the prime number theorem, $\pi(2^{w+1})-\pi(2^w)\asymp2^w/(w+1)$. The finitely many remaining intervals are absorbed by a single constant using nonemptiness. Unique factorization and the disjoint intervals make a tuple of lower primes recoverable from its product. Hence its atom probability is at most

$$2^{-S_L}\prod_{j\le L}C(w_j+1) =2^{-S_L+O(L\log(L+1))}.$$

Conditioning on the finitely many higher coordinates leaves that law unchanged. The source-shell count uses no short-interval theorem: the discarded interval has at most $X_{k-1}$ integers, and $X_{k-1}\ln X_k/X_k\to0$.

A sufficiently large deterministic $k_0$ ensures that a source prime in shell $k$ exceeds every auxiliary prime through $k$. The exponent interval $[q_j+1,2q_j-1]$ contains exactly $q_j-1$ representatives modulo $q_j-1$, so the discrete-log digit exists uniquely. All digits are below their radices. Mixed-radix uniqueness and $X_k<Q_k$ prove injectivity of the labels within a shell; the last nonzero digit distinguishes different shells.

For sums, the largest digit is $4q_j-2<4q_j$, so no carries occur. The double, single, and zero digit ranges are disjoint. They recover both shell indices from a sum. A same-shell equality gives congruent products modulo $Q_k$; both products are below $Q_k$, since $2\gamma<1$. Unique factorization then identifies the unordered pair, including a diagonal.

For unequal shells, only the implication from a sum equality to a fiber of $(P\bmod N,PR\bmod M)$ is needed. Fixing $P$ permits cancellation modulo $M$ and uses $Y<M$. Fixing $R$ permits cancellation modulo $M$, then combines the two coprime moduli and uses $X<MN$. There is no cancellation of a lower-shell prime modulo $N$. Thus both fiber projections are injective. Distinct points have nonzero upper and product differences: shell disjointness excludes a crosswise factorization. Their sizes force $N<X$ and $M<XY$, and consequently

$$1-\gamma<\lambda=\frac{S_L}{S_K} <\frac{\gamma}{1-\gamma}.$$

In particular all potentially contributing lower shells are comparable in scale to the upper shell. No unproved L-claim is needed for this transfer.

Weakest step 1: degeneracies in the fixed-modulus count

I rederived the two equations used for elimination from

$$P_1(R_2-R_1)+NaR_2=Ms,\qquad P_1(R_3-R_1)+NbR_3=Mt.$$

Eliminating $P_1$ gives $NA=MB$, where

$$A=aR_2(R_3-R_1)-bR_3(R_2-R_1),\quad B=s(R_3-R_1)-t(R_2-R_1).$$

Since $(M,N)=1$, $A=dM$ and $B=dN$ for an integer $d$. The direct triangle bound is $|d|<2XY^2/(MN)=W$.

For $\lambda\le2/3$, the first three comparisons in (14) have a negative fixed multiple of $S_K$ plus $O(R_K)$ in their logarithms. The logarithm of $WU^3/M$ is at most

$$1+(4\gamma-4)S_K+(2\gamma+2)S_L \le 1+\frac{16\gamma-8}{3}S_K,$$

whose coefficient is negative. This verifies all four comparisons uniformly, including lower endpoints of the source shells.

If $d=0$, reducing $A=0$ modulo the distinct lower prime $R_2$ gives $R_2\mid bR_3R_1$, and hence $R_2\mid b$. But $0<b<U<R_2$. This rules out the first degeneracy without a sign assumption on the product offsets.

For $\Delta=at-bs\ne0$, put $\rho=R_2-R_3$, $\sigma=atR_2-bsR_3$. Multiplying the first equation in (13) by $t-s$ and substituting the second gives

$$\rho\sigma=d\{M(s-t)-NaR_2+NbR_3\}.$$

The inverse linear transformation is

$$R_2=(\sigma-bs\rho)/\Delta,\qquad R_3=(\sigma-at\rho)/\Delta.$$

Substitution and multiplication by $\Delta$ give exactly

$$(\Delta\rho-dN(b-a))(\Delta\sigma-dNab(s-t)) =d(s-t)\{\Delta^2M+dN^2ab(b-a)\}.$$

This verifies both the signs and the factor of $s-t$ in (16).

The right side cannot vanish. Here $d\ne0$, $s\ne t$, and vanishing of the braces would imply $M\mid dab(b-a)$, by coprimality with $N$. Yet $0<|dab(b-a)|<WU^3<M$. This also excludes cancellation when $d$ is negative.

Every parameter and factor in this formula has magnitude $2^{O(S_K)}$, uniformly over outcomes and lower shells. The exposed divisor argument is valid: primes at most $\sqrt{\ln H}$ contribute at most $O(\sqrt{\ln H}\log\ln H)$ to $\log\tau(n)$; larger primes have total multiplicity at most $2\ln H/\log\ln H$. The first contribution is also $O(\ln H/\log\ln H)$ eventually. Thus there are $2^{O(S_K/\log S_K)}$ signed ordered factor pairs. Each determines $\rho,\sigma,R_2,R_3$, then $R_1$ from $s\ne t$, and $P_1$ from $R_2\ne R_1$. The remaining upper coordinates follow. Invalid reconstructions only reduce the count.

For the separate case $\Delta=0$, coprimality after dividing $a,b$ by their gcd gives $a=ea'$, $b=eb'$, $s=fa'$, $t=fb'$, with $f\ne0$. Writing $C_0=fM-NeR_1$ gives

$$a'C_0=P_2(R_2-R_1),\qquad b'C_0=P_3(R_3-R_1).$$

The inequalities $a'<P_2$, $b'<P_3$ force both distinct upper primes to divide $C_0$. Also $|C_0|<2XY<P_2P_3$. Thus $C_0=0$, contradicting distinct lower coordinates. This handles the branch where the inverse linear transformation would otherwise be unavailable.

Finally, $U^2$, $4Z^2$, and $2W$ bound the numbers of gaps, signed offsets, and nonzero $d$'s. Their product has main factor

$$U^2Z^2W=\frac{2X^5Y^4}{(MN)^3}.$$

This gives (18), including its error scale and uniformity. The larger divisor error need not fit inside $O(R_K)$: in this part of the band a fixed exponential saving absorbs it later.

Weakest step 2: the line contribution and conditional averaging

I checked the lattice bound with the line contribution retained. For a primitive integer normal, the two proposed kernel vectors have cross product $c$. The scalar product of $c$ with the proposed third vector is one, so the three vectors form an integer basis of determinant one. This proves both the kernel basis and its covolume $\|c\|_2$. The special normal with its first two coordinates zero is immediate.

A shortest nonzero lattice vector is primitive in lattice coordinates and extends to a basis. Reducing the other basis vector along it makes its parallel component at most half the shortest length. Minimality then forces the perpendicular height to be at least $\sqrt3/2$ times that length. Counting parallel lattice lines and points on each line inside a ball containing the cube yields

$$O(1+Y/\ell+Y^2/D).$$

An area term alone would not justify the subsequent argument.

After reducing $s,t$ by their gcd, their values and $s-t$ are all nonzero. A prime outside the three upper primes cannot divide the entire normal $((t-s)P_1,-tP_2,sP_3)$, since it would divide both reduced offsets. A common prime equal to one of the upper primes would have to divide a nonzero reduced offset multiplying a different upper prime. That offset has smaller magnitude than the putative common prime, which is impossible. Thus the normal is primitive, with length at least $P_-h$.

For a kernel vector $v$, the first brace in

$$P_1\{(t-s)v_1-tv_2+sv_3\} +N\{-ta v_2+sb v_3\}=0$$

is divisible by $N$. If $\|v\|_\infty<N/(3h)$, its absolute value is below $N$, so both braces vanish. The transformed integer vector

$$((t-s)v_1,-tv_2,sv_3)$$

is then a nonzero integer multiple of $((b-a)/g,-b/g,a/g)$. Its second coordinate implies $\|v\|_\infty\ge b/(gh)$. Consequently every nonzero vector has length at least $\min\{N/(3h),b/(gh)\}$, which gives (21). All cancellations here use $P_1$ coprime to $N$.

The small-coefficient hypothesis is justified uniformly:

$$\log_2(Z/P_-)=-(1-\gamma)S_L+\gamma w_K.$$

The band supplies $S_L>(1-\gamma)S_K$, while $w_K=o(S_K)$. Thus even the unreduced difference of the offsets is eventually below every upper prime.

Under conditioning on the higher auxiliary primes, $N$, the source shells, the candidate upper coordinates, and the ranges $U,Z$ are fixed. For each candidate point and nonzero $s$, at most one lower product can work:

$$M=(P_2R_2-P_1R_1)/s.$$

It has probability zero if inadmissible, and at most $2^{-S_L+O(R_K)}$ otherwise. Repeated candidate points producing the same $M$ do not require independent indicators: finite linearity of expectation gives an upper bound including all their multiplicities.

The gcd sum is also uniform. For a fixed positive $b$,

$$\sum_{0<a<b}\gcd(a,b)\le b\tau(b),$$

by summing common divisors. Summing $\tau(b)$ by divisors gives $\sum_{b<U}\tau(b)\le U(1+\ln U)$. The signed offset sums are bounded respectively by $O(Z^2)$, $O(Z^3)$, $O(Z^2)$ for the weights $1,h,1/h$.

Multiplying these counts by the three terms in (21) yields, before division by $M_0$, precisely

$$XU^2Z^2,\quad XYUZ^3,\quad XYU^2Z^3/N,\quad U^2Y^2Z^2,$$

up to $2^{O(R_K)}$. The logarithm in the gcd sum and the factor $X/P_-=2^{\gamma w_K}$ fit inside that error. Using the lower bound for $N$, their coefficients after division by $M_0$ are

$$\begin{array}{c|l} 1&5\gamma-2+(2\gamma-1)\lambda\\ 2&5\gamma-1+(4\gamma-3)\lambda\\ 3&6\gamma-3+(4\gamma-1)\lambda\\ 4&4\gamma-2+(4\gamma-1)\lambda . \end{array}$$

The first minus fourth is $\gamma(1-2\lambda)<0$; the third minus fourth is $2\gamma-1<0$. This independently recovers (25)--(27), with the line term still present.

Weakest step 3: critical endpoints, one outcome, and every cutoff

Let $\eta=3-20\gamma/3$. Its positivity follows from $\sqrt6<49/20$, using $6<(49/20)^2$. The deterministic coefficient on $\lambda\le2/3$ is at most $\gamma-\eta$. Its uniform $o(S_K)$ divisor error is eventually smaller than $\eta S_K/2$.

On the remaining band, the line coefficient $H_\gamma(\lambda)=5\gamma-1+(4\gamma-3)\lambda$ decreases, and $H_\gamma(2/3)=\gamma-\eta$. The area coefficient $D_\gamma(\lambda)=4\gamma-2+(4\gamma-1)\lambda$ increases, and

$$D_\gamma\!\left(\frac{\gamma}{1-\gamma}\right) =\frac{5\gamma-2}{1-\gamma}=\gamma,$$

because $\gamma^2+4\gamma-2=0$. This is a non-strict endpoint bound. The proof correctly uses no positive distance from that endpoint. Summing fewer than $K$ lower shells therefore yields $\mathbb E T_K\le2^{\gamma S_K+C_1R_K}$ with a single constant.

Each $T_K$ is a finite-cylinder, nonnegative finite random variable. Markov's inequality with threshold $2^{\gamma S_K+C_1R_K+K}$ makes its failure probability at most $2^{-K}$. For any $J$, the union of failures after $J$ has probability at most $\sum_{K\ge J}2^{-K}$, which tends to zero. Thus a single infinite auxiliary sequence has only finitely many failures. This uses no independence between shell events.

For a finite three-uniform hypergraph, independently retaining vertices with probability $p$ gives expected vertices minus edges $pn-p^3m$. An outcome at least as good as the expectation exists. Deleting at most one vertex per surviving edge leaves an independent set. The two specified cases of $p$ give a fixed positive constant times $n/\sqrt{1+m/n}$. Combining this with $n_K=2^{\gamma S_K-O(\log S_K)}$ and $m_K\le2^{\gamma S_K+O(R_K)}$ gives (32).

All lower shells used in defining an upper hypergraph are full source shells. Hence the upper independent sets may be selected simultaneously: later or earlier deletions cannot create an unrecorded triple. Lexicographic choice of a maximum independent set makes each finite selection definite; exceptional early shells are empty. If three distinct representations remained, digit ranges would force a common shell pair. The equal-shell case was already excluded. In the unequal case, fiber injectivity supplies three distinct selected upper vertices, contradicting independence. This covers all integer sums and diagonal conventions.

For every sufficiently large real $x$, there is a unique $K$ with $B_K\le x<B_{K+1}$. All selected shell-$K$ labels are below $x$, and with $t=\log_2x$,

$$S_K+2K<t<S_K+w_{K+1}+3K+3.$$

Therefore $t-S_K=O(R_K)$ and $R_K=O(\sqrt{t\log_2 t})$. The shell lower bound gives

$$\log_2 A(x)\ge\gamma t-O(\sqrt{t\log_2 t}).$$

Multiplying by $\ln2$ changes the loss to a constant times $\sqrt{\ln x\,\ln\ln x}$, since $\ln(\log_2x)=\ln\ln x-\ln\ln2$. Constants absorb this fixed logarithm-base correction eventually. The finitely many exceptional shells only enlarge $x_0$. Since $R_K=o(S_K)$, the retained shell sizes tend to infinity. The loss divided by $\ln x$ tends to zero, which proves the same-set consequence for every fixed $\alpha<\gamma$.

Strongest attempted refutation

The main attack was to seek a positive-dimensional family missed by the fixed-modulus divisor count. Two dangerous possibilities are $\Delta=0$, where the coordinate recovery is singular, and $\Delta\ne0$ with the right side of (16) zero, especially for negative $d$. Either would invalidate the finite divisor-pair count and could destroy the saving needed below $\lambda=2/3$.

The attack failed under the actual hypotheses. In the singular branch, both distinct upper primes divide $C_0$, while $|C_0|<P_2P_3$, forcing equal lower coordinates. In the resonant branch, coprimality forces the nonzero integer $dab(b-a)$ to be a multiple of $M$, while its magnitude is below $M$. These contradictions do not discard negative offsets or assume generic parameters.

I also tested the reason a line term is essential. In a potential arithmetic-progression configuration $a=u,b=2u,s=1,t=2$, with an integer $u>0$, the collision plane has the short direction $(1,1,1)$. Its diagonal integer points would defeat an area-only estimate in suitable scales; they are not asserted to be actual fiber triples with distinct lower primes. The proof retains and sums the line contribution, whose exponent $H_\gamma$ remains strictly below $\gamma$ on the upper band. Finally I allowed shell ratios arbitrarily close to the critical upper endpoint. The argument still closes with a non-strict main exponent and the explicit $O(R_K)$ loss.

Audit checklist

Exact premise scope and source reading

No native L-claim, local component theorem, or prior independent report is consumed as an established premise. L4 is recorded as tier 0 in the admitted index; its old loss and scope match the comparison. Its theorem is not applied to the new schedule. The subject supplies the necessary digital, factorization, lattice, averaging, and selection arguments, all reviewed here. There is no batch acceptance order or unproved mathematical antecedent of the asserted existence theorem.

Ordinary foundational arithmetic used includes unique prime factorization, Bézout's identity and gcd divisibility, finite linearity of expectation, and elementary limits and inequalities. Their relevant applications and cancellation hypotheses are exposed above. The nontrivial standard external interfaces and reading boundaries are:

• Erdős, Beweis eines Satzes von Tschebyschef, admitted 1932 artifact: section 5, printed 197--198, PDF pages 4--5, both visually inspected. The interface is existence of a prime $m<q\le2m$ for each integer $m\ge1$, applied at $m=2^{w_j}$. The inclusive upper endpoint and the $m=1$ case agree with the visible source. The theorem statement and application were checked; the full external proof was not independently reconstructed.
• Conrad, Cyclicity of the nonzero residues modulo a prime: Theorem 2.6, PDF/printed page 3, visually inspected. The interface is cyclicity of the multiplicative group of each prime field, applied separately at every selected $q_j$. The downloaded SHA256 exactly matches the admission's recorded hash. The statement was checked; the displayed local proof was read without recursively reviewing its earlier lemmas.
• Zagier, Newman's Short Proof of the Prime Number Theorem, 1997: printed 705, PDF page 2, visually inspected. The interface is $\pi(x)\sim x/\ln x$. It supplies the dyadic auxiliary-prime count and the source-shell count by the explicit endpoint deductions above. The theorem statement was checked, not its analytic proof.
• Durrett, Probability: Theory and Examples, version 5, 11 January 2019: the version was visually verified on PDF page 1. Theorem 2.1.21, printed 53/PDF 61, supplies a probability measure for consistent finite-dimensional real laws; products of the finite uniform prime laws satisfy consistency and concentrate on the intended prime sequences. Theorem 1.6.4, printed 29--30/PDF 37--38, gives Markov after taking $\phi(u)=\max(u,0)$ and a positive tail threshold. Theorem 2.3.1, printed 68/PDF 76, requires summable event probabilities and no independence. All these pages were visually inspected. Statements were checked; the two short displayed probability arguments were read, and the needed Markov and tail-union deductions were independently rederived. The extension theorem's proof was not reviewed.

The Cilleruelo manuscript was not opened. Its method attribution and the subject author's claimed manuscript reading are not certified. L2 was not opened, so its historical/current standing assertion is not independently authenticated here. Neither supplies a density premise or an essential omitted deduction. No fresh literature census, priority assessment, public acceptance investigation, or exact catalog-page status audit was commissioned or performed.

Exact external artifact identities

The original source-access date was 15 September 2026. These are the reviewer's reported pinned versions, not sources newly acquired or visually inspected during native filing. The URLs, theorem locators, specializations and actual reading depth are given in the preceding source list.

The nine full-page visual inspections were Erdős PDF pages 4 and 5; Conrad page 3; Zagier page 2; and Durrett pages 1, 37, 38, 61 and 76. The source record additionally counted five Erdős and four Conrad extracted-text page surfaces and one Durrett title-page text extraction: nineteen page surfaces in that conservative count, not nineteen full visual inspections. Some automatic web text was partial. Requested URLs and successful responses were recorded, but redirect chains were not separately logged.

Resource and acceptance limits

The review used only the frozen mathematical subject and the narrow local comparison described above. It acquired the four identified PDFs and visually inspected the nine selected complete pages. Public source proof coverage stops at the interfaces and displayed portions explicitly listed. No mathematical code is required to reproduce the deductions. Source-view rendering was a reading aid, not computational reproduction of the theorem.

The original source record measured 3,270,316 downloaded PDF bytes and 1,664,312 bytes of derived page images. Its conservative repeated-PDF accounting charged 26,758,522 bytes; adding a text/metadata allowance gave 27,807,098 bytes. Raw web-tool transfer bytes were unavailable and were not claimed as measured. Each web open was charged the subsequently pinned artifact size instead. Repeated instruction reads and administrative allowances were likewise conservatively accounted, not silently treated as single reads. These are the original review's reported accounting limitations, not measurements repeated during filing.

The report is a mathematical assessment of the pinned subject only. Changing its mathematics or its consumed premises requires a new scope assessment. This report does not grade its own independence or contract compliance, confer a tier, or by itself establish ordinary-clone integration. The distinct grade and this author-prepared native retention remain separate from the historical mathematical verdict.