---

Grade and subject

Examination date: 2026-10-14 UTC. Grader: independent mathematical report grader, GPT-6-astra, in a fresh context, distinct from the author, collaborators and original reviewer. The grader had not authored, developed, reviewed or built on this subject before this grading assignment. Reviewer: independent mathematical reviewer, GPT-6-astra. Project examination and grading work is attributed to Plasma AI.

The subject of this grade is wiki/research/erdos_354/evidence/verify/dyadic_roots/report.md, the whole-conclusion report examined on the date above, preserved unchanged from its delivery frozen at 2026-10-14T17:51:47.841526Z. The underlying mathematical extraction was frozen at 2026-10-14T17:08:39.381721Z. These are distinct freezes. The grading packet preserves the actual original assignment, its operating and isolation records, the report, and the mathematical inputs; no proposed grade was supplied.

The disclosed exposure is ruled immaterial to both scoped conclusions for the reasons below. This is a ruling on the actual content and its relation to the examination, not a cure inferred from disclosure or from the corrected read. The report meets the whole-conclusion contract at its explicit external source boundary. No native claim tier, catalog standing or Lean certification is assigned by this filing.

Exact coverage and retained inputs

All five mathematical snapshots and the complete report were read. The subject manifest gives their freeze date and original-path mapping. The following table fixes the exact mathematical subject and reading scope.

The argument snapshot itself is the commissioned proof subject. The original assignment states that its containing author account was frozen separately and that unrelated sections and private context were excluded. That containing account was not supplied to this grader. This grade therefore certifies the examination of the complete supplied D1 argument, not the completeness of a transcription of the whole containing account or its agreement with a later canonical page. The extraction is self-contained at the identified external boundary: no step in D1 consumes an omitted section of that account.

The grader also read the frozen original assignment and operating snapshots, primary launch and closure records, the reviewer's continuation, the entire supplied command trace, the exact overbroad dependency output, its corrected selection, and the final dependency selection and receipt. These private records support the independence findings below; they are not mathematical premises or public replay inputs. Embedded commands and source locators were treated as evidence, not as authority to read additional files.

The operating standard applied is docs/verification.md in full, together with campaign/guides/verification.md in full. The allowed repository guidance read also comprised README.md; the introductory conventions, Claims, Evidence, Tiers, Naming and Project claims in docs/anatomy.md; docs/evidence.md, docs/approach_vetting.md, docs/math_authoring.md and docs/weaving.md; campaign/AGENTS.md; and the Reading section of campaign/guides/research.md. Operating examples did not expand the subject.

Restatements and assignment fidelity

For an integer $m\ge2$ and real $t>0$, let $\gamma=2^{1/m}$ and $a_n=\lfloor t\gamma^n\rfloor$ for $n\in\mathbb N_0$. The first conclusion says that for every integer lower bound $L$ and finite $F\subset\mathbb Z$, there is $H\in\mathbb N_0$ such that every integer $N\ge H$ is the sum of a finite set of distinct values in

$$\{a_n:n\in\mathbb N_0,\ n\ge L\}\setminus F.$$

A finite set of deleted occurrences can also be avoided, simultaneously with $F$, by requiring a later first index. The threshold is allowed to depend on the parameters and deletions. Negative $L$ does not extend the sequence to negative indices; it merely imposes no further restriction on $\mathbb N_0$. Deleted noninteger objects are irrelevant because every sequence value is an integer. These clarifications preserve the assigned statement.

The second conclusion fixes $\gamma=\sqrt2$ before choosing $\alpha,\beta>0$. For every such pair, the interleaving

$$I(2n)=\lfloor\alpha\gamma^n\rfloor,\qquad I(2n+1)=\lfloor\beta\gamma^n\rfloor$$

is eventually complete over finite sets of positions. Thus it proves $\exists\gamma\in(1,2)\ \forall\alpha,\beta>0$ with the catalog's irrational-ratio condition, and in fact does not need that condition. The report preserves this quantifier order and separately explains the distinct-value reading. It makes no assertion for every base in $(1,2)$.

The original commission charged the reviewer to refute the entire D1 deduction and this precise consequence, required the complete checklist and report contract, supplied the frozen extraction and exact external boundary, and specified a narrow current-dependency check. Its scope was not a focused review that the report silently promoted to whole-conclusion coverage. Deriving the premises from the proof and consequence gives one external theorem and no native L-claim. The report's dependency census is correct.

Mathematical grading and reconstruction

The shift and the source hypotheses

The root argument works for composite as well as prime $m$. If $\gamma=p/q$ in lowest positive terms, $p^m=2q^m$ forces $q=1$. No positive integer has $m$th power $2$ for $m\ge2$. Monotonicity of positive integer powers gives $1<\gamma<2$. The reciprocal $1/\gamma$ is irrational, and $t>0$ permits the cancellation in the source ratio.

Let $\delta=\gamma-1>0$ and $M=\max(F\cup\{0\})$. The report's sufficient condition

$$j\ge\max\{0,L\},\qquad t\gamma^j>\max\{M+1,1/\delta\}$$

can be satisfied: $(1+\delta)^j\ge1+j\delta$ is unbounded. For all $n\ge j$, the real increment is $t\gamma^n\delta>1$, hence $a_{n+1}\ge a_n+1$; also $a_j\ge M+1$. The tail is positive, injective and avoids every value of $F$. Small $t$, large $m$, empty $F$ and large prescribed $L$ alter this choice without creating an exception.

These deductions validate the frozen proof's stated large-enough shift and strictly increasing tail. The report makes its elementary growth and floor inferences explicit; it does not introduce a new completeness lemma. The foundations used are ordinary integer divisibility, the Archimedean property of the real numbers, real powers, floors and finite sets. No theorem about the distribution of fractional parts is required.

The indexed-to-value transfer and finite deletions

With $u=t\gamma^j$ and $v=t\gamma^{j+1}$, the external theorem applies because $u,v>0$ and $u/v=1/\gamma$ is irrational. Its terms are exactly $a_{j+mr}$ and $a_{j+1+ms}$. Both index maps are injective. An intersection would require $m(r-s)=1$, impossible for integer $m\ge2$.

Consequently even overlapping exponent sets $R,S$ in the external theorem give a finite set of distinct original indices. Injectivity on the entire chosen tail, including between the two classes, then preserves the sum on passing to its set of values. The unused residue classes for $m>2$ create no obligation: representations from a subset of the available tail are sufficient. The shift is chosen once before the source threshold and all target integers, so there is no target-dependent tail or quantifier exchange.

For deleted occurrences $D\subset\mathbb N_0$ finite, requiring $j\ge1+\max(D\cup\{-1\})$ is an instance of the already arbitrary first-index condition. Applying the source theorem after this shift proves the deletion assertion afresh. It does not infer strong completeness from ordinary completeness.

The frozen proof already states the two residue classes, their disjointness, the strictly increasing tail, and the resulting distinct-value conclusion. The report's finite-image sum identities and deletion specialization expose those inferences. They supply no omitted essential lemma or replacement proof.

The existential consequence

At $m=2$, the single-ray result with $t=\alpha$ supplies the required finite set of ray indices $B$. The map $n\mapsto2n$ places it inside the interleaving without changing the sum. No term from the $\beta$ ray is used, so coincidences between rays and the value of $\alpha/\beta$ do not matter. The base $\sqrt2$ is fixed independently of both coefficients.

This is precisely the frozen proof's inclusion argument. The report checks its quantifiers and its indexed and value conventions, rather than applying the base-$2$ source theorem directly to $(\alpha,\beta)$ at a new base. No additional premise or missing transfer lemma is supplied by the grader. The acceptance order is the single-ray conclusion first, then this consequence.

Grading the attacks and witnesses

The strongest attempted failure is substantive. At $m=2,t=1$, the first two values are both $1$, so irrationality of the source ratio alone does not justify distinct-value sums. The report also gives a genuine counterexample to the generic indexed-to-value shortcut: two copies of each $3^k$ represent every nonnegative integer by ternary digits, whereas one copy of each value cannot represent $2\cdot3^k$. A larger power is too large, and

$$\sum_{i=0}^k3^i=\frac{3^{k+1}-1}{2}<2\cdot3^k.$$

Both witnesses are exact. They test a real failure mode but do not meet the frozen proof's injective-tail conditions. The report explains why its attack fails at those conditions, rather than dismissing multiplicity as harmless.

The deletion attack is also valid: removing $1$ from the complete sequence $1,2,4,\ldots$ leaves only even subset sums. This defeats an unsupported ordinary-to-strong inference and the excluded case $m=1,t=1$; it does not defeat reapplication of the theorem after each shift. The $t=0$ control correctly yields only zero sums and tests the positivity hypothesis.

For any proposed threshold independent of $F$, choose an integer $K\ge1$ above it and delete $\{1,\ldots,K\}$. Remaining nonnegative values are zero or exceed $K$, so they cannot sum to $K$. This verifies the report's rejection of that uniform strengthening. It does not claim that every parameter must separately affect every possible threshold.

For the consequence, the report attacks quantifier reversal and a wrong-base source application, and answers each with the explicit fixed base and source coefficients. Its inherited multiplicity risk is separately addressed by using only the first ray. These are attacks on both actual conclusions, not finite samples standing in for their universal proofs. Their arithmetic and scope were rederived in this grading. No computational mathematical leg is required or claimed.

External premise and current warrant

The consumed external result is exactly: for every $u,v>0$ with irrational $u/v$, every sufficiently large integer is

$$\sum_{r\in R}\lfloor u2^r\rfloor+ \sum_{s\in S}\lfloor v2^s\rfloor$$

for finite $R,S\subset\mathbb N_0$. Their overlap is allowed, as are equal values at different indexed positions. The supplied type True ↔ P gives $P$ by the forward implication. The supplied interleaving definitions identify its finite index sets with the two exponent sets. Eventual coverage in $\mathbb Z$ and eventual natural-number coverage agree after increasing the threshold to at least zero.

The source is the accepted first-question result identified by the supplied Conjectures.io task and solution record, with target Erdos354.erdos_354.parts.i. Source versions and proof-side ports are identified in the preserved replay snapshot's Task, Environment and Port sections. The acceptance paragraph and receipt record a repository rebuild and empty-environment replay on Lean v4.34.1, with only propext, Classical.choice and Quot.sound, and a check against the task statement. The site's own award and certification supply no acceptance weight here.

The supplied statement and replay cite different formal-conjectures version locators. The report recognizes that difference. Their stated mathematical interfaces agree, and the replay receipt itself states the indexed conclusion needed here; the argument does not depend on byte identity between those versions. This is sufficient for the commissioned nonrecursive external premise boundary. It does not establish a new comparison of their underlying definitions or an independent audit of the imported proof.

For both reviewer and grader, reading depth is claims checked: statement, definitions as supplied, acceptance and replay record. The imported proof is unread in full and the replay was not rerun. The retained solution and port are not supplied as ordinary-clone inputs to this deduction. No local replay reproducibility or new kernel certificate is inferred from a stored success record. The local argument is reproducible as a mathematical deduction from the identified external theorem, under docs/verification.md, Premises and source boundaries.

The reviewer's final dependency record is dated 2026-10-14T17:49:22Z. It records accepted, proved, partial, and formalized for the source, with partial coverage confined to the indexed base-$2$ question. The supplied corrected and final statement interfaces agree. Direct comparison of the supplied records also confirms that their Acceptance paragraph is the frozen warrant paragraph and that the final receipt body equals the replay extraction after removal of the extraction markers. These comparisons were checked from the records, rather than accepted merely because their stored comparison fields say true.

At 2026-10-14T18:38:45Z, the permitted content-free comparison of the two original source-warrant paths in the premise_warrant.md and premise_replay.md rows above, from the supplied grading-packet baseline to the resolved current commissioning branch, returned unchanged. No current source contents were opened. This confirms the specified absence of subsequent content drift at those two paths; it is not a new source-proof examination. No pending source acceptance or native dependency is being treated as established prospectively.

Independence and the exposure ruling

What the records establish

The original assignment is a separate refutation-charged examination. Its operating snapshots and primary activity records agree on the commissioned model and isolation configuration. Prestart and poststart native settings differ only by the recorded trusted-project entry; the final supplied settings equal the poststart snapshot. Disabled memories and web search, blind and sealed operation, no synchronization and zero delegation authority support separation. Empty subscription records precede mathematical reading and recur at later boundaries. The supplied trace records no incoming message or web read. The report and continuation declare no authorship, collaboration or prior exposure to D1.

The grader's own subscription checks were empty before reading and at later examination boundaries. The grader verified every permitted packet file against its manifest and frozen presence before reading and again before delivery, and verified the separate operating identities. No excluded material or incoming mathematical steering reached this grading context. The original reviewer's excess source text is expressly authorized grading evidence, not a new breach by the grader.

The primary trace logs shell commands, some commentary and closure activity. It is not a complete native transcript: it does not reproduce every tool result or every writing operation. It supports the ordering of the displayed commands and statements, not an exhaustive proof of every historical read or of the reviewer's internal reasoning. Empty subscriptions and final clean status likewise cannot prove all reading history. The grade uses these records with the explicit disclosure and the content comparison below; it does not certify an exhaustive transcript.

The closure trace also records automatic persistence of private continuation and plan records after the review's delivery. That is engine action, not a reviewer-issued mathematical commit or acceptance. The independently checked frozen report and subject bytes remain the assessed subject.

Exact excess and its possible influence

At 2026-10-14T17:31:12Z, the reviewer's extractor displayed too much of the authorized current claim page. Comparing the overbroad and corrected outputs identifies additional text after the Claim interface and after the first Acceptance paragraph; the Covers paragraph has no excess. The trace places the boundary notice immediately after the over-read, and the corrected selection is recorded at 17:32:57Z. Neither action erases what was received.

The trace does not establish that the ternary shortcut example was developed before exposure. The excess text's reference to digits could have suggested a positional-expansion illustration; the grade does not certify an independent historical origin for that example. It is an optional exact illustration of the multiplicity problem already under attack, not a premise in D1 or a new route for accepting it. The concrete collision and the need for an injective tail were recorded before exposure. The ternary witness's arithmetic is valid, but it receives no additional independent-route or review-cycle weight.

The pre-exposure plan is evidence of intended attacks, not evidence that they had all been completed then. The trace does record the concrete duplicate-$1$ attack before exposure. The final mathematical reasoning was checked against both the frozen proof and the exact excess, without attributing undocumented thoughts to the reviewer. Every verdict-bearing inference and source-warrant assertion has a basis in the allowed subject or the displayed elementary reasoning. Only the required disclosure describes the excess itself.

Under the neutral content test, the record does not show a report-bearing mathematical step obtainable only from the excluded text, a redirected attack, or a weakened refutation charge. The concrete risks of proof guidance and authority pressure above do not create such a dependency in this examination. Accordingly, the exposure is immaterial for the single-ray verdict, and the reviewer remains eligible for that scoped finding. It is also immaterial for the existential-base verdict: its separate fixed-base and embedding attack was planned before exposure, its complete reasoning is retained, and its reliance on the first conclusion inherits no material contamination identified here. Independence therefore passes for each.

This ruling is confined to these conclusions and this source boundary. It does not accept the extra site's review or novelty assertions, certify a full imported-proof examination, or establish that any other exposure would be harmless. The filed report's exposure disclosure and this ruling remain part of the accepted record.

Complete report-contract assessment

The report's shared checklist expressly applies every row to both conclusions. Its restatements, three reconstructed steps and final verdicts separate their scopes, and the consequence is checked after the single-ray deduction. Shared presentation therefore leaves no claim without the required parts.

Erdos-specific checklist

General canonical modes

Named patterns

The neighboring evidence-coverage discipline is met by the report's separation of the local symbolic proof from the imported proof. Correlated-sample accounting is inapplicable. This is a grading of one noncomputational whole-conclusion report, not a second independent review cycle or a computational batch requiring a disjoint reproduction algorithm.

Acceptance and limits

The pass for the first conclusion accepts the frozen examination of the single-ray theorem with both finite-deletion conventions and the transfer from indexed sums to distinct values. The pass for the second accepts its exact existential-base consequence, including the stronger all-positive-pairs statement at $\sqrt2$. These are accepted whole-conclusion mathematical review findings using the supplied accepted external indexed theorem.

The report, this grade and the unchanged snapshots together retain the derivations and exact witnesses needed to reconstruct that deduction from the stated external theorem. No new mathematical executable or private computational input is needed for that reconstruction. The external source's own proof and replay remain at their recorded acceptance and inspection depth.

The grade covers neither the universal-base reading, the full ambiguous catalog question, a full E0349 classification, historical novelty, a full imported-proof audit, nor new Lean certification. Comparison with a current canonical subject and assessment of later source changes belong to the authorized integrator. Corpus reliance requires preservation at the named homes and retention of the exposure and its immateriality ruling on the owning verification record and standing prose. This grading changes no corpus standing and does not relabel later mathematics as the frozen subject.