---

Subject and independence

Role. Independent reviewer in a fresh context, given only the commissioning assignment. The reviewer took no part in writing the connections page, the reconstruction pages, the lead pages or any filed review or grade of this folder, and opened none of the folder's filed reviews or grades.

Subject. The section "No further consumer" of the Problem 354 connections page, wiki/research/erdos_354/connections.md as it stood at 2026-09-28T13:50:48Z, lines 611--652: the H2 heading at line 611 through the last line of the file. The section holds no proof displays. Its content is one interface summary of the descent, one recognition rule, one inventory of ten catalog problems with descriptors, one obstruction sentence carrying two mathematical reasons, two corpus-coverage sentences and an outcome line.

Lead pages. The section's "two leads named above" are rational_ratio_dyadic_floor_descent and small_base_two_ray_distance_divergence, both read in full as of the same time.

Reading and depth. All as of 2026-09-28T13:50:48Z. The section in full. The page's preamble (lines 20--57: purpose and standing, the definition of $A_{\alpha,\beta}$, complete and strongly complete sets, the normalized pair and its weights, conversions, events, prefix sums, moduli and the run invariant) as shared input. Reconstruction pages as premises: the Theorem 5.1 page, the normalization page and the theorem page in full; the definitions and statements of the erosion, mesh and projection lemma pages; the statements only of the Remark 4.1, Remark 4.2, Theorem 9, Corollary 12 and (BG) pages. Problem pages, statement paragraphs only: 343, 344, 345, 346, 347, 351, 253, 246, 123, 125, and, as the targets the section refers to, 354 (statement and its formulation paragraph), 349, 348, 254 and 124. Guidance: the "Whole-claim report", "Audit checklist" and "Grading and claim standing" sections of docs/verification.md, docs/approach_vetting.md in full, the "Source fidelity" section of docs/evidence.md and docs/math_authoring.md in full. No library PDF was opened: the section invokes no source theorem directly; its only imported mathematics is the reconstruction pages, which the assignment makes premises.

Exposures. Five, none used in a verdict. (1) The status line of the Problem 347 page ("Proved (Lean)") was displayed while the statement paragraph was located; the other fourteen problem pages were cut before their status lines. (2) The two lead pages, which are allowed reading, paraphrase the Problem 354 page's Current assessment in their "Target and status evidence" sections. (3) The H2 headings and the one-line "Outcome" sentences of the page's other six sections were seen while locating the targets and leads the section refers to; no other line of those sections was read, and none of their reviews was opened. (4) A file-name listing of wiki/theory/ and wiki/research/ (paths only, no content) was taken to check the coverage sentence S7. (5) The listing of the folder's evidence/verify/ directory shows that reviews of the other sections exist; none was opened. The reviewer also ran a short exact-integer computation for the attack on S5; its derivation is retained below and no run log is kept.

Restatement

Conventions carried from the page: for $\alpha,\beta>0$, $A_{\alpha,\beta}$ is the set of nonzero floors $\lfloor2^n\alpha\rfloor$, $\lfloor2^n\beta\rfloor$, $n\ge0$; a normalized pair has $N=\lfloor\beta\rfloor<M=\lfloor\alpha\rfloor<2N$, $N\ge2$, weights $a_i=\lfloor2^i\alpha\rfloor$, $b_i=\lfloor2^i\beta\rfloor$, conversions $u_i=a_{i+1}-2a_i$, $v_i=b_{i+1}-2b_i$, prefix sums $P_n$ (subset sums of the $2n$ weights of indices below $n$), $D_n=\gcd(a_n,b_n)$ and $h_n=h(P_n\bmod D_n)$, the longest run of missing residues. A set is complete if every sufficiently large integer is a sum of distinct elements, strongly complete if it stays complete after every finite deletion.

• S1 (interface of the descent). The folder's reusable mechanism is the Theorem 5.1 descent: for a normalized pair, if at layer $n$ the weights are $a_n=dp$, $b_n=dq$ with $d=D_n$, $\gcd(p,q)=1$, the conversions at indices $n,\ldots,n+\ell-2$ vanish, the conversion at index $n+\ell-1$ is nonzero, and $K=2^\ell\ge K_*(p,q)$ with $K_*(p,q)=2q(p-1)+4(p+q)+64$, then $h_t\le\max(0,h_n-1)$ for every $t\ge n+\ell+3$ and every choice of later conversions; the connection of the mesh nodes is a finite certificate valid for all $q<p<2q$; the inputs are the interlacing $b_i<a_i<2b_i\le b_{i+1}$, the base-2 digit recurrence $u_i,v_i\in\{0,1\}$, and the reduced ratio at the block's first layer.
• S2 (recognition rule). A consumer of the descent is a set equal, up to finitely many elements, to a union of two rays $\lfloor2^n\alpha\rfloor$, $\lfloor2^n\beta\rfloor$, or a question about subset sums of such a set.
• S3 (inventory). Beyond the targets of the earlier sections (354, 349, 348, 254, 124), the corpus problems on complete sequences, subset sums of distinct terms, floors of geometric sequences and dyadic intervals are 343 and 344 (subcompleteness under a density condition), 345 (thresholds of completeness of the powers $n^k$, open), 346 (strongly complete sequences that lose completeness under every infinite deletion), 347 (a sequence with ratio tending to $2$ whose cofinite subsequences have subset sums of density $1$), 351 (strong completeness of $p(n)+1/n$), 253, 246 and 123 (sums of distinct $a^kb^l$ and $a^kb^lc^m$) and 125; all but 345 are resolved.
• S4 (two-ray sets and Problem 347). No two-ray set with ratio not a power of $2$ is an example for Problem 347, because the consecutive ratios of its sorted terms alternate near $\alpha/\beta$ and $2\beta/\alpha$ (normalized values) and so do not tend to $2$.
• S5 (two-ray sets and Problem 346). No two-ray set is an example for Problem 346, for the same reason.
• S6 (single rays). A single ray $\lfloor2^n\alpha\rfloor$ has ratios tending to $2$ and at most $2^n$ subset sums below $a_n\ge\lfloor\alpha\rfloor2^n$; together with the resolution of both problems this excludes single rays from Problems 346 and 347.
• S7 (coverage). The theory tree holds no claim on complete sequences, and no research lead other than the folder's two concerns them.
• S8 (outcome). No second consumer exists, so no method page is made; the two leads are the folder's only routes.

Checklist

• Quantifiers and scope. One failure and one gap. S5 quantifies over all two-ray sets while its reason bears only on the conclusion of Problem 346, not on its hypotheses (F1). S6's displayed bound proves the intended density statement only when $\lfloor\alpha\rfloor\ge2$ (F2). In S1, "every later layer" means every $t\ge n+\ell+3$, the bound is relative to $h_n$, and no layer-to-layer monotonicity is asserted (F4). S4 is correctly quantified: the normalization removes finitely many elements and finitely many coincident values, neither of which changes the limit points of the ratio sequence.
• Circularity. None. The section derives nothing from its own conclusion; the descent is consumed only through the Theorem 5.1 page.
• Model and convention changes. Checked; no unproved transfer. Problem 346 asks for a strictly increasing set and Problem 347 for a nondecreasing sequence; a normalized pair has pairwise distinct values (item 4 of the normalization page), and the finitely many pre-normalization coincidences contribute finitely many ratios equal to $1$. "Complete" in Problem 346 is defined in its statement exactly as on the page. The ratios "near $\alpha/\beta$ and $2\beta/\alpha$" are those of the normalized coefficients, as the page's notation paragraph makes standing; for an unnormalized pair they are the normalized ratio $\theta\in(1,2)$ and $2/\theta$.
• Finite and statistical overreach. None in the section. The reviewer's witness below uses an exact finite check only as the base of an induction whose step is proved by an explicit inequality.
• Uniformity. Pass. The dependence of the descent constant on the reduced ratio is displayed as $K_*(p,q)$; the certificate lemma of the Theorem 5.1 page is stated for all integers $q<p<2q$, all twelve templates and all four third-digit pairs, so "uniform" is warranted.
• Extremal conclusions. Inapplicable; the section states no infimum, supremum or sharpness. The count bound of S6 was checked in its own units (number of subset sums against $a_n$).
• Consequences and composition. "None consumes the descent" composes the descriptors of S3 with the obstructions S4--S6. The 347 link and the single-ray link hold (S4, S6 with F2); the 346 link is defective (F1) and, if Problem 346 is resolved in the negative, the composition still holds because a resolved problem needs no consumer; that dependence on the resolution is not visible in the sentence. "So no method page is made" is a convention consequence outside the read set.
• Computation. Inapplicable to the section, which runs nothing. The reviewer's computation is exact integer arithmetic (integer square roots and a bitset of subset sums) and is retained as a derivation.
• Reproduction. Inapplicable; the section states no rerun command or coverage claim.
• Source and verdict fidelity. The ten descriptors agree with the statement paragraphs of the problem pages (Problem 351's set is recorded there as $\{p(n)+1/n\}$, as the section writes). The interface summary agrees with the Theorem 5.1 page up to F4. The status words ("proved", "solved", "disproved", "open") are outside the read set and were not checked. The section claims author-recorded standing only, as the page's preamble states.

Weakest steps

W1: the ratio pattern (S4, S5). Item 4 of the normalization page gives $b_i<a_i<2b_i\le b_{i+1}$, so the merged sorted list of a normalized pair is $b_0<a_0<b_1<a_1<\cdots$ and its consecutive ratios are $a_i/b_i$ and $b_{i+1}/a_i$. Since the floor error is below $1$, $2^{-i}a_i\to\alpha$ and $2^{-i}b_i\to\beta$, hence $a_i/b_i\to\theta=\alpha/\beta\in(1,2)$ and $b_{i+1}/a_i\to2/\theta\in(1,2)$. The set of limit points of the ratio sequence is $\{\theta,2/\theta\}$; it never contains $2$, and it is a single point exactly when $\theta=\sqrt2$, in which case the ratio converges to $\sqrt2$. For Problem 347, whose sequence must satisfy $\lim a_{n+1}/a_n=2$, this excludes every pair whose ratio is not a power of $2$: normalization multiplies each coefficient by a power of $2$, which replaces each ray by a tail, so the unnormalized set differs by finitely many elements and finitely many coincident values, and neither affects the limit points. This is S4. For Problem 346 the hypotheses are (i) $A\setminus B$ complete for every finite $B$, (ii) $A\setminus B$ not complete for every infinite $B$, (iii) $a_{n+1}/a_n\ge1+\varepsilon$ for all $n$, and the question is whether the ratio tends to the golden ratio. The pattern shows that (iii) holds (every ratio exceeds $1$ and both limit points exceed $1$, so the infimum over $n$ is above $1$) and that the questioned conclusion fails; it says nothing about (ii). So the pattern does not establish S5; see the attack below.

W2: the single-ray count (S6). For a ray with coefficient $\alpha>0$, the elements below $a_n=\lfloor2^n\alpha\rfloor$ are among $a_0,\ldots,a_{n-1}$ (the sequence is nondecreasing), so the subset sums below $a_n$ number at most $2^n$, and $a_n\ge\lfloor\alpha\rfloor2^n$ because $\lfloor\alpha\rfloor2^n$ is an integer not exceeding $2^n\alpha$. Both displayed facts are correct. The lower density of the subset sums is therefore at most $1/\lfloor\alpha\rfloor$, which is below $1$ exactly when $\alpha\ge2$. The sharper bound $a_n>2^n\alpha-1$ gives lower density at most $1/\alpha$, below $1$ whenever $\alpha>1$. Discarding the zero terms puts the coefficient in $[1,2)$; the coefficient $1$ gives the powers of $2$, whose cofinite subsequences obtained by deleting $r\ge1$ elements have subset sums of density $2^{-r}$ (the integers with zero binary digits at the deleted positions). Hence no single ray is an example for Problem 347, and no single ray is strongly complete as Problem 346 requires: for a coefficient in $(1,2)$ infinitely many integers are unrepresented, and for the powers of $2$ one deletion leaves infinitely many unrepresented. The conclusion of S6 holds for every ray; the displayed inequality proves it only for $\lfloor\alpha\rfloor\ge2$.

W3: the interface (S1). The Theorem 5.1 page's hypotheses of Section 3 are exactly the layer data $a_n=dp$, $b_n=dq$, $\gcd(p,q)=1$, the zero conversions at $n,\ldots,n+\ell-2$, the nonzero conversion at $n+\ell-1$, and $K=2^\ell\ge K_*$. Its proof uses, beyond the certificate table and the three finite lemmas: the interlacing $q<p<2q$ at layer $n$; the digit bounds $u_i,v_i\in\{0,1\}$ (for the constant part at most $22$ and for $b_r\le8dKq+15$); the prefix bound $S_n<L_n=d(p+q)$, which the normalization page derives from the recurrence and positive initial terms; $0\in P_n$ (so $k\le d$); and the doubling bound on all later weights in sorted order, which is the interlacing at every later layer. Each of these is an instance of "the interlacing" or "the digit recurrence" of a normalized pair, so S1 names the inputs at their actual strength. The conclusion is $h_t\le\max(0,h_n-1)$ for $t\ge n+\ell+3$, uniform in the later conversions; the page's Scope note disclaims monotonicity. S1's "every later layer" compresses this (F4) but changes no quantifier that the folder relies on.

Strongest attack

S1. Sought a hypothesis of the descent absent from S1's list, by listing every fact the Theorem 5.1 proof consumes (W3). All reduce to the interlacing and the digit recurrence of a normalized pair plus the finite certificate, which S1 names. Sought a strengthened quantifier: S1 does not claim monotone descent, and its "every later layer" reads as the permanent bound once the offset three is supplied. The attack failed.

S3. Compared each descriptor with the statement paragraph of its problem page. Problem 343's density condition is $\gg N$, Problem 344's is $\gg N^{1/2}$, both asking for subcompleteness; Problem 345 compares thresholds of completeness of consecutive powers $n^k$; Problem 346 and 347 are as restated in W1; Problem 351 is recorded as strong completeness of $\{p(n)+1/n\}$; Problems 246 and 123 are sums of distinct $a^kb^l$ and $a^kb^lc^m$; Problem 125 concerns the lower density of a sum of two digit-restricted sets; Problem 253 is an implication for sequences with ratio tending to $1$. Every descriptor is faithful. The status words are outside the read set. The attack failed on what could be read.

S4. Sought a two-ray set whose sorted ratios tend to $2$. By W1 the limit points are $\theta$ and $2/\theta$, both in $(1,2)$; a pair whose ratio is a power of $2$ is a single ray as a set, and as a multiset its doubled values give ratios $1$ and about $2$ in alternation, so it has no limit either. The attack failed.

S5. The attack succeeded against the argument and left the statement unresolved. The reason offered, the ratio pattern, concerns the conclusion of Problem 346 and not its hypothesis (ii). A witness that (ii), not the pattern, decides membership: the normalized pair $\alpha=3\sqrt2$, $\beta=3$ ($N=3<M=4<6$, ratio $\sqrt2$ irrational), whose sorted ratios converge to $\sqrt2$, so a fragile such set would have been a counterexample to the questioned implication rather than a non-example. Its weights are $a_n=\lfloor2^n\cdot3\sqrt2\rfloor$, the integer square root of $18\cdot4^n$, and $b_n=3\cdot2^n$: the weights of indices below $3$ are $3,4,6,8,12,16$, and $a_3=33$ (since $33^2=1089\le1152<1156=34^2$), $b_3=24$. An exact enumeration shows that the subset sums of $\{3,4,6,8,12,16\}$ contain every integer of $[6,43]$ (and not $5$ or $44$). Delete the infinite set $B=\{b_m:m\ge4,\ 4\mid m\}=\{48,768,\ldots\}$ and add the retained weights of indices $\ge3$ in increasing order to $W_0=[6,43]$, whose span is $37$. The mesh lemma with $k=1$ keeps a full interval and adds the weight to the span whenever the weight is at most the current span. Let $s_m$ be the span before the weights of index $m$ are added; then $s_m\ge37+\sum_{i=3}^{m-1}a_i+\sum b_i$, the second sum over retained $3\le i<m$. Since $a_{i+1}\ge2a_i$, $\sum_{i=3}^{m-1}a_i\ge33(2^{m-3}-1)$, while $a_m\le33\cdot2^{m-3}+2^{m-3}-1$ because $a_{i+1}\le2a_i+1$, and $b_m=24\cdot2^{m-3}$. Hence $s_m\ge b_m$ for every $m\ge3$, and $s_m-a_m\ge5-2^{m-3}+\sum b_i$ over retained $3\le i<m$; one of $m-1,m-2$ is retained for $m\ge4$, contributing at least $3\cdot2^{m-2}>2^{m-3}$ for $m\ge5$ and $24>2$ for $m=4$, and $s_3=37\ge33=a_3$. So every integer $\ge6$ is a sum of distinct elements of $A_{3\sqrt2,3}\setminus B$: the pair violates (ii) and is not an example for Problem 346, by an argument unrelated to its ratios. The same construction applies to any pair for which some $P_r$ contains a full interval longer than $a_r$, but the section proves nothing of the kind for all pairs, and the descent yields strong completeness, not the deletion structure, so the general statement S5 stays unresolved within the read set. If Problem 346 was resolved by proving the implication, S5 follows from that theorem and W1; the section does not say so.

S6. Sought a coefficient for which the displayed bound is vacuous: $\alpha=3/2$ has $a_n=3\cdot2^{n-1}$ for $n\ge1$, $\lfloor\alpha\rfloor=1$, and the bound reads $a_n\ge2^n$, giving density at most $1$; the sharper bound gives at most $2/3$ (F2). The conclusion survives for every ray by W2; the attack succeeded only against the displayed inequality's sufficiency.

S7. Scanned the file names of the theory tree (claims L1 to L22 on covering systems, additive bases, powerful complements, unit distances, extremal graphs, roots of unity, totients, dissociated sets and rainbow cycles) and of the research leads; none names complete sequences other than the folder's two leads, which are filed under research/leads/, outside the folder's directory (F5). A page could concern complete sequences under another name; the read set does not allow opening them. The attack found nothing.

S8. Consistent with the outcome lines of the other sections (one transfer and one conditional transfer, each recorded as a lead, three failures, and one transfer to Problem 254); whether the Problem 254 transfer is a "route" in the section's sense could not be checked without reading that section. No refutation.

Premises

• Theorem 5.1 page: hypotheses of Section 3, statement $h_t\le\max(0,h_n-1)$ for $t\ge n+\ell+3$, the certificate lemma, the length condition (5.2) and the Scope note; read in full. Independently reviewed as of 2026-09-28T05:03:27Z with corrections applied since; treated as established at that standing.
• Normalization page: items 1--6 and the reduction; used for the interlacing, the distinct values, the digit bounds, the prefix bound $S_n<L_n$ and the fact that normalization takes tails; read in full; same standing.
• Theorem page: strong completeness of $A_{\alpha,\beta}$ for irrational ratio, used only to note that a two-ray set with irrational ratio meets hypothesis (i) of Problem 346; read in full; same standing.
• Mesh lemma page: Lemma 2.2 and its consequence, used in the S5 witness; definitions and statement read; same standing. The erosion and projection lemma pages were read for definitions only.
• Statements of the Remark 4.1, Remark 4.2, Theorem 9, Corollary 12 and (BG) pages, read to assess F3; not otherwise used.
• The two lead pages: author-recorded and unreviewed by their own "Current review" sections; used only to identify the leads and the mechanisms the page reused (F3).
• Problem statements as listed under Reading; no external source and no PDF.
• No batch acceptance order applies; the review is focused and stands alone.

Findings

• F1 (required). Location, lines 641--643: "the two-ray sets are not examples for Problem 346 or 347, since their sorted terms have ratios alternating near $\alpha/\beta$ and $2\beta/\alpha$". Defect: for Problem 346 the ratio pattern bears on the questioned conclusion (a ratio tending to the golden ratio), not on the hypotheses (strong completeness, loss of completeness under every infinite deletion, ratios bounded away from $1$). A two-ray set with irrational ratio meets the first and third by the folder's theorem and the pattern itself, so the pattern cannot exclude it; had such a set been fragile under every infinite deletion it would have been a counterexample. The obstruction therefore does not block what it claims for Problem 346. Witness: the pair $(3\sqrt2,3)$ of the attack on S5, whose ratios converge to $\sqrt2$ and which fails the deletion hypothesis by an explicit infinite deletion. Proposed replacement: "the two-ray sets are not examples for Problem 347, whose sequence must have ratio tending to $2$, since their sorted terms have ratios with limit points $\alpha/\beta$ and $2\beta/\alpha$; Problem 346 asks about sequences that lose completeness under every infinite deletion, a property the descent does not address, and a two-ray set with that property would be a counterexample rather than an example, so nothing is transferred either way and the problem is resolved". If Problem 346 was resolved by proving its implication, write instead that the exclusion follows from that theorem because the ratios of a two-ray set do not tend to the golden ratio.
• F2 (suggested). Location, lines 644--645: "a single ray, whose ratios tend to $2$, has at most $2^n$ subset sums below $a_n\ge\lfloor\alpha\rfloor2^n$". Defect: the displayed bound gives a subset-sum density below $1$ only when $\lfloor\alpha\rfloor\ge2$; for a coefficient in $(0,2)$ it is vacuous. Witness: $\alpha=3/2$ (W2, S6 attack). Proposed replacement: "has at most $2^n$ subset sums below $a_n>2^n\alpha-1$, so once its zero terms are discarded, which puts $\alpha$ in $[1,2)$, its subset sums have lower density at most $1/\alpha<1$ unless $\alpha=1$, when the ray is the powers of $2$ and deleting $r\ge1$ elements leaves subset sums of density $2^{-r}$".
• F3 (note). Location, line 613: "The reusable content of the folder is a descent". The folder also reconstructs the all-even floor mechanism at Pisot and Salem bases (the Theorem 9 and Corollary 12 pages) and the one-per-dyadic-interval example (the Remark 4.1 page), and the lead pages record that this page reused both (Proposition 354.7; the transfer to Problem 254). The recognition rule of the next sentence is tailored to the descent, while the search terms ("floors of geometric sequences", "dyadic intervals") do cover the other mechanisms, so the inventory is not narrowed. Suggested wording: "The reusable content of the folder not consumed above is a descent".
• F4 (note). Location, lines 614--615: "lowers the modular run invariant of every later layer". Precisely, $h_t\le\max(0,h_n-1)$ for every $t\ge n+\ell+3$, relative to the invariant at the block's first layer; the layers $n+1,\ldots,n+\ell+2$ are not covered and no monotonicity from layer to layer is asserted. Suggested wording: "of every layer from three past the block".
• F5 (note). Location, lines 648--649: "no research lead outside this folder concerns them". The two leads are filed under research/leads/, outside research/erdos_354/; the sentence is true only in the reading "no lead other than the two named above". Suggested wording to that effect. The check was by file names only.

Verdict

• S1: refutation-failed. The interface names the descent's inputs at their actual strength; F4 is a wording note.
• S2: not a claim; a recognition heuristic.
• S3: refutation-failed for the content descriptors, each faithful to its statement paragraph; the status words were not checkable within the read set.
• S4: refutation-failed. The limit points of the sorted ratios are the normalized ratio and its complement to $2$, never $2$.
• S5: proof defective; statement unresolved. The ratio pattern does not bear on the deletion hypothesis of Problem 346; the witness pair shows membership is decided by that hypothesis, and the general statement is not established by the section (F1).
• S6: refutation-failed for the displayed bound; the exclusion it is meant to carry holds for every ray by the sharper bound, which the section should state (F2).
• S7: refutation-failed on a names-only check, with the wording note F5.
• S8: not a claim; an outcome record consistent with what was read.

Limitations: the status words of the ten inventoried problems, the direction in which Problem 346 was resolved, the convention that a method page needs two consumers, and the content of the other sections of the page were outside the read set; the coverage sentence S7 was checked by file names only; the S5 witness settles one pair and any pair with a finite full-interval certificate, not the general statement. No tier is warranted or withheld by this review; every statement of the section remains author-recorded. This focused review assigns no tier and changes no status.