# Futer–Wise: source audit and reviewer evidence

**Review scope:** arXiv:2106.04497v3, the supplied desktop report, and selected passages of the published 2024 edition.

The desktop report identifies a genuine invalid inference in the self-overlap proof. It persists in the journal edition. Neither the counting lemma nor the main random-quotient theorems have been refuted by these checks.

The central source check is independently reproducible. A full replacement self-overlap counting proof is not supplied. These are AI-assisted assessments and source checks, not independent human peer review or formal verification.

[Reviewed manuscript](paper.pdf) · [Published edition](journal-2024.pdf) · [Original desktop report, 3 November 2026](desktop-report.pdf)

## Self-overlap

In the unit-edge Cayley tree of $F(a,b)$, set $g=a^m b^{3m}$, with $m\geq4$, and let $L$ be its axis through $1$. The maximal $J=1$ loose piece and its companion on $aL$ are

$$s=[1,a^m b],\qquad s'=[a,a^m],\qquad |s|=m+1,\quad |s'|=m-1.$$

For $h_N=ag^N$ with $N\geq2$, the axes remain fixed while

$$h_N^{-1}s'=g^{-N}[1,a^{m-1}],\qquad s\cap h_N^{-1}s'=\varnothing.$$

Nevertheless the projection of $h_N^{-1}s'$ to the parameter circle lies inside the projection of $s$. For $m=4,N=2$, the lifted intervals have coordinates $[0,5]$ and $[-32,-29]$; modulo the period $16$, the second becomes $[0,3]$.

The axes diverge at the ends of their common interval and cannot rejoin in a tree. The first excluded vertices are distance two from the other axis, establishing maximality. The cyclic word is not a proper power. Taking $d=m$ and $\ell=4m$ gives $d>2J$, $d\leq\ell/3$, and $|s|<\ell/2$. Arbitrarily large $m$ meets every fixed logarithmic lower threshold. Subdividing doubles all lengths and $J$.

[Exact check script](verify_report.py) · [Fifteen finite checks and rounding certificate](report-check-results.json)

**The self-overlap objection is valid.** Demonstrated error in a proof step; counting lemma not refuted.

**Source:** Definition 4.8; Lemma 4.11(2), arXiv p.20 / journal p.642; Lemma 4.13.

Replacing h by h g^N preserves the translated axis and the two occurrences on the relator circle, but translates the pulled-back companion by whole periods. The chosen lifted intersection can therefore be empty while the circular intersection is nonempty. The preceding normalization puts the start of s near a basepoint; it does not fix this freedom. The asserted disjointness, and hence the proposed fundamental-domain decomposition, does not follow. The same step remains in the journal text.

**Consequence:** This is a central obligation in the supplied proof of the self-piece probability bound. Its downstream use in the main theorems remains conditional on fixing that count. It is not a counterexample to either main theorem, nor to the numerical probability bound.

**Remaining work:** Normalize the representative before splitting into small and large overlap, handle circular wrapping and truncation, and give the resulting finite encoding. The desktop report offers a plausible repair route, not a completed proof.

## Cocompactness

**Separate properness from cocompactness.** Properness bridge located; full scope of Theorem 3.5 remains unresolved.

**Source:** Theorem 3.5 and Figure 1, journal p.629; end of proof, p.634.

The theorem as printed does not assume the original group is hyperbolic. Its proof attributes cocompactness to compactness of X without supplying the bridge. The cited Wise statements were located: Remark 5.46 supplies metric properness for bounded cone diameters; Corollary 5.30 supplies quasi-isometric wall carriers; Proposition 7.4 supplies cocompactness in the hyperbolic, quasiconvex-wall setting. These are statement/hypothesis checks, not independent reproving of the book.

**Consequence:** For each fixed finite presentation, compact relators give the required uniform diameter bound. In the main hyperbolic applications, Lemma 3.4 and the standard wall-stabilizer argument support the report's proposed cocompactness route. This does not prove the broader standalone formulation. No counterexample to that formulation is established.

**Remaining work:** Supply the scope-appropriate cocompactness argument or add its missing hypothesis. The midpoint wall should explicitly cross the geodesic once, with its antipodal partner excluded; the existing strict inequalities leave room for a mesh adjustment. The final freeness conclusion should name the dual cube complex.

## Zero-growth

**The zero-growth-exponent constant needs correction.** Confirmed local error; main density loss is avoidable.

**Source:** Proposition 4.7, equation (4.2), arXiv pp.15–16 / journal pp.636–637.

The final denominator 1−exp(−a) is zero at a=0. Retaining the previous denominator 1−exp(−[aα+b(1−α)]) fixes that division for fixed α<1 and b>0. For α<1/20 it has the uniform positive lower bound 1−exp(−19b/20). This correction alone does not justify the earlier pure exponential hyperplane-neighborhood count: infinite cyclic groups have linearly growing balls, not bounded balls.

**Consequence:** A bounded-width annulus count, or an a+ε upper bound chosen inside the strict density margin, addresses the elementary-stabilizer issue without changing the limiting main density. A complete rewritten counting proof was not supplied in this audit.

**Remaining work:** Correct both the displayed constant and the use of subgroup growth; do not claim that fixing the denominator alone completes the proof.

## Growth-metric

**The preliminary growth statement is too broad.** Counterexample to the stated preliminary metric-space assertion.

**Source:** Definition 2.2 and Theorem 2.3, arXiv pp.5–6 / journal p.627.

Give F₂ the left-invariant metric d(g,h)=sqrt(|g⁻¹h|). The triangle inequality follows from sqrt(x+y)≤sqrt(x)+sqrt(y); balls are finite and the left action is transitive, hence proper and cocompact. Its integer-radius-n ball has exactly 2·3^(n²)−1 elements. Thus finite generation and proper cocompactness on an arbitrary metric space do not imply exponential orbit growth. Using F₃ and its free factor F₂ also defeats the claimed finite strict growth-exponent comparison.

**Consequence:** The cited Dahmani–Futer–Wise theorem uses a graph; the Matsuzaki–Yabuki–Jaerisch setting is proper, geodesic and hyperbolic. The main geometric applications have the additional structure, so this example does not refute them.

**Remaining work:** State the geometric hypotheses needed for the imported growth results.

## Essential-core

**The essential-core equality is stronger than needed.** Supported correction to the stated reduction.

**Source:** Definition near Theorem 1.1; main proofs, arXiv pp.25,31 / journal pp.647,653.

The paper calls a hyperplane essential when the whole space is not in a bounded neighborhood of it; Caprace–Sageev require both halfspaces to be deep. In the report's rank-three rose with a collar on a rank-two subrose, the collar mid-hyperplane has stabilizer F₂ and satisfies the paper's weaker condition. The standard core removes that shallow collar. The ambient exponent remains log 5, while the maximal hyperplane-stabilizer exponent drops from log 3 to 0. Product with a finite interval triggers the paper's conditional pruning step.

**Consequence:** The safe statement is b_core=b and a_core≤a. Keeping the original a as an upper bound preserves the advertised density restriction. The example challenges equality, not the main theorem.

**Remaining work:** Use the standard essentiality convention consistently and replace the unnecessary equality for a by the inequality.

## Relator-identity

**Equality of relator lifts needs its indexing convention.** Supported clarification request; no main-theorem refutation.

**Source:** Definition 3.2, arXiv p.7 / journal p.628.

Equality of image subsets can hide coincidence of elevations of differently indexed relators. The author-linked earlier Wise preprint, Section 3.c, explicitly retains the alternative i≠j in its cone-piece definition. This supports the need to retain relator identity. The later Arenas reference was not used as an input to the original proof.

**Consequence:** The interpretation must be fixed before applying small cancellation. The report correctly presents this as a convention issue.

**Remaining work:** Specify indexed elevations or explain the relator identifications being imposed.

## Local-constants

**The local constant and range corrections check out.** Local corrections; no demonstrated change to the final thresholds.

**Source:** Lemma 4.13; Lemma 4.15; Proposition 5.7; introductory density discussion.

The union of five bad events requires the sum of their bounds, not just the maximum. Replacing J by J+2δ requires an appropriate common overlap threshold. Proposition 4.14 need only be used at d=2qℓ≤ℓ/3, after which monotonicity controls larger pieces. In Proposition 5.7 the total δ coefficient is −4, while the printed simplification has −2. The 1/41 endpoint is not covered by a strict density hypothesis.

**Consequence:** These repairs adjust constants, a large-length threshold, or wording. They do not themselves repair the central self-overlap count.

**Remaining work:** Apply the local corrections while keeping the central proof obligation separate.

## Density

**The quantitative density arithmetic is consistent.** Checked conditional calculation.

**Source:** Main density formulas and the pentagonal quasi-isometry formula.

Balancing 2q=α(1−q) gives q=α/(α+2). At the limiting α=1/20 this is 1/41; the wall restriction gives (b−a)/20. The quasi-isometry conversion gives 20λ and 40λ+1. Independent evaluation gives λ≈1.5627191146324275 and 40λ+1≈63.5087645852971. An exact rational Taylor-bound certificate also proves 40λ+1<63.51, with a positive margin greater than 9471971/7671972100.

**Consequence:** The rounded denominator is in the safe direction. This calculation is conditional on the overlap estimates and the geometric quasi-isometry theorem; it is not a verification of every figure-dependent argument in Section 6.

**Remaining work:** Retain the strict inequalities and the report's conditional language.

## Axis-date

**One blind-review concern is resolved by an earlier source.** Chronology uncertainty resolved.

**Source:** Haglund, arXiv:0705.3386v1, Theorem 1.4 and Theorem 6.3.

The OpenAI assessment flagged uncertainty because the manuscript bibliography lists a 2023 publication. The relevant theorem and the subdivision condition are already in the 23 May 2007 preprint. For the deck action, nontrivial elements cannot fix a vertex, so the axis alternative applies after subdivision.

**Consequence:** This is not an outstanding chronological defect in the paper. It illustrates why the raw model findings must be checked rather than copied into a verdict.

**Remaining work:** Cite the dated preprint when documenting the original proof inputs.

## Coverage and costs

The campaign delivered one complete full-paper OpenAI assessment and one complete Gemini assessment of Lemma 4.11 with six source contexts. Three earlier full-paper attempts and the final focused Anthropic arbitration reached their output caps. The final arbitration returned truncated text; it supplies no completed adjudication. All six attempts are closed.

The authorized campaign budget was $10. The conservative usage estimate, including all incomplete attempts, is $5.6468. This is a usage estimate at frozen batch rates, not a provider invoice.

| Model | Scope | Effort | Result | Estimated usage |
|---|---|---|---|---:|
| gpt-6.1-sol | blind-panel | high | Completed | $0.3631 |
| claude-opus-5-5 | blind-panel | max | Output cap; incomplete | $1.1664 |
| gemini-4-argon | blind-panel | high | Output cap; incomplete | $0.4478 |
| claude-fable-5-1 | report-audit | max | Output cap; incomplete | $2.4386 |
| gemini-4-argon | focused-gemini-v2/wave-000 | high | Completed | $0.2948 |
| claude-fable-5-1 | focused-arbitration | high | Output cap; incomplete | $0.9360 |

[OpenAI assessment](openai-assessment.json) · [Focused Gemini assessment](gemini-assessment.json) · [Coverage and usage record](coverage-and-cost.json)

The focused Gemini response guessed three printed numbers incorrectly: its intended references are Theorem 2.4, Remark 2.5, and Definition 4.8. The quotations and native labels identify the intended passages. Its additional warm-up-count and uniform-range concerns remain unadjudicated.

## Limits and source record

- Detailed coordinator audit concentrates on the supplied report and the central dependency path, not every argument in the paper.
- Native full-text models received the entire expanded TeX, but no figure images. Failed or capped deliveries supply no verdict.
- Selected journal passages were compared and critical pages visually inspected; this is not a line-by-line collation of both complete editions.
- External theorem statements and selected applicability hypotheses were inspected; their full proofs were not independently reproved.
- The 2021 Wise numbering was checked in a third-party transcription of a dated typeset proof; the older author-linked preprint corroborates the cocompactness criterion with different numbering.
- The proposed normalization repair, complete cyclic counting, and the sharp figure-dependent pentagonal lower bound remain outside the completed proof checks.
- Additional OpenAI observations, including the literal scope of Corollary 1.4 and some encoding details, remain in its raw assessment and were not automatically promoted to confirmed defects.
- The focused Gemini warm-up-count and uniform-range objections remain unadjudicated. Fable returned a truncated response, so the native arbitration is incomplete and supplies no validated verdict.

[External-reference inspection record](reference-checks.json) · [Targeted edition comparison](version-comparison.json) · [Source versions and hashes](source-editions.json) · [Structured coordinator findings](coordinator-findings.json)

No author response was supplied. No message was sent to the authors. Publication does not change these coverage or theorem-status limits.
