{"id":"16ecf0ca-e58b-4d6f-92f3-6658fdd199be","arxiv_id":"2607.15905","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A mesh-CSG method that builds the arrangement once with exact integer predicates and recovers ambiguous radial orders by weighted-majority votes over topological relations, yielding fast N-ary booleans.","lead":"This paper presents trueform, a mesh-CSG pipeline that keeps only topology exact and resolves disagreements between rounded geometric readings by majority vote within topological units. It reports watertight boolean outputs 5.5–100× faster than prior libraries and interactive in the browser.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exactness is conditional on an unproven majority: correlated materialisation flips from shared vertices/edges can push a relation past 50%, so 'exact topology' is not a guarantee.","rationale":"The paper's central claim is that topological aggregation recovers the intended answer exactly via majority vote. The reader correctly identifies the weakest assumption: the majority must carry the intended order. The paper is unusually candid, stating in §5.1 that the guarantee is statistical and that an adversary can defeat it. My stress-test does not introduce a new objection; it sharpens the same concern by noting that materialisation errors are not independent categorical noise, as Appendix A idealises. Shared vertices and edges create correlated flips, so a single rounding event can affect multiple observations in a relation and potentially cross the 50% threshold. The empirical perturbation sweeps in §4.1 provide strong evidence that the vote works on the tested distributions and at the tested error magnitudes, and the shipped implementation is a real artifact. However, the abstract's unqualified 'exact' and 'recovered' claims are broader than the proven conditional guarantee. The reader's CONDITIONAL verdict, requiring narrowed language and prominent statement of adversarial limits, is appropriate. My concern does not move the verdict; it reinforces the need for those changes. A concrete, reproducible failure test would settle whether the majority assumption can actually be broken in a realistic snap pattern, which would either validate the current claim or demonstrate the overstatement.","tokens_in":29621,"tokens_out":8481,"duration_ms":90307,"concrete_test":"Reproduce the two-sphere sweep of §4.1 but force correlated failure: choose a sphere offset d and a coarse integer grid so that a single latitude vertex ring's snap causes orient3d to flip for a contiguous majority of edges in the intersection circle's relation. Run trueform's build and compare the extracted domain partition against the exact partition from the original planes. If the partition differs (or the output boolean's signed volume is wrong), the majority vote fails under a realistic snap pattern, confirming the exactness claim is conditional.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee of topological aggregation is definitional (Appendix A): the vote returns the intended permutation exactly when a majority of the relation's observations carry it. The paper's empirical sweeps (§4.1) show the vote tolerates two-to-three orders more construction error than per-edge reading, but they do not establish that real materialisation error is unbiased or that a majority failure is unreachable. Adjacent edges in a relation share endpoints, so a single rounded vertex can flip several radial-order observations at once; under near-coplanar geometry these flips can be systematic rather than independent. §5.1 concedes 'a configuration can be built to defeat it.' Thus the abstract's 'we keep it exact' overstates what is proven: for decisions spanning faces, the method returns a maximum-a-posteriori estimate, not a worst-case exact answer. If a correlated rounding flips >50% of a relation's weighted observations, the output topology (and hence the boolean) is wrong, not merely geometrically imprecise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents trueform, a mesh-CSG pipeline that computes arrangements with integer-exact predicates on snapped coordinates, classifies all contacts into five canonical types, arranges each cut face locally in its own plane, and uses a two-level identity scheme to keep the arrangement consistent without global exact constructions. Around non-manifold edges, radial order is decided by exact predicates on the original input planes. Where a decision spans multiple faces, the method aggregates possibly disagreeing per-edge radial-order observations by a weighted-majority vote within each topological relation, casts this as maximum-a-posteriori estimation, and classifies volumetric domains by per-domain inclusion bitvectors. One arrangement build serves arbitrary boolean expressions of any arity, including open surfaces declared as sheets. The paper reports robustness experiments (two-sphere sweeps, a geological arrangement, Thingi10K, perturbation sweeps), performance scaling experiments, and comparisons against six prior libraries plus browser baselines.","tokens_in":29899,"tokens_out":6737,"duration_ms":105783,"significance":"If the claims hold, the paper makes a strong practical contribution: a fast, header-only, cross-platform CSG library that builds one arrangement and answers a whole family of booleans, handles non-manifold input and open sheets, and still uses exact predicates without exact constructions for its in-build radial structure. The benchmark protocol is clearly specified, based on public data (Thingi10K) and external implementations, with no fitted constants, and the code is released. The limitations section is unusually candid, explicitly acknowledging that the majority vote is statistical and that materialised output can self-intersect. However, the exactness claim is conditional on a majority assumption whose failure modes are acknowledged but not bounded, and the validity oracle used to grade outputs is weaker than the paper's wording suggests. These issues are load-bearing for the paper's central 'exact topology' claim and need to be resolved before acceptance.","major_comments":[{"comment":"The abstract and Contribution 3 state that the intended answer is 'recovered' by topological aggregation and that 'we keep it exact.' But the guarantee is definitional: Appendix A Eq. (4) identifies the MAP estimate with the weighted majority, and the appendix states that 'the vote returns the intended permutation exactly when a majority of the relation's observations carry it.' Section 5.1 concedes that a configuration can be built to defeat the majority. The §4.1 experiments show that the vote tolerates more error than per-edge reading under the tested noise model, but they do not establish that real materialisation errors are unbiased, independent, or bounded below the 50% threshold in the worst cases described in Figure 2 (a single rounded vertex can flip several radial-order observations at once). The face-spanning decision is therefore a maximum-a-posteriori estimate with a conditi","section":"Abstract, §1.2, §5.1, Appendix A (Eq. 4)"},{"comment":"The validity oracle grades only closure, consistent orientation, signed volume, and surface area. Section 5.1 admits that materialising the output can introduce self-intersections that the exact arrangement did not have. Therefore the paper's claim of 'valid, watertight output' (abstract) and the validity counts in Tables 5 and 7 do not establish that the output is a valid solid in the usual CAD sense; a watertight but self-intersecting mesh is not a valid solid for downstream stages. Either extend the oracle to check self-intersection (e.g., by exact arrangement of the output or at least by reporting crossing counts on a representative subset), or explicitly define 'valid' as 'closed and consistently oriented' and state that self-intersections may remain.","section":"§4.3, 'Validity' and §5.1"},{"comment":"The perturbation experiment models construction error by perturbing non-manifold-edge endpoints independently in a uniform box. This does not exercise the correlated failure mode the paper itself identifies in Figure 2 and §5.1: when adjacent edges share an endpoint, a single rounded vertex can flip several radial-order observations at once, and near-coplanar geometry can make such flips systematic rather than independent. The claimed margin of 'two to three orders of magnitude more construction error' is therefore established for a particular independent-noise model, not for the worst relevant materialisation process. The paper should add a correlated-corruption experiment (e.g., coherently jittering shared vertices, or constructing a near-coplanar configuration tuned to approach the majority threshold) to show how close realistic inputs come to the failure regime, or provide a theoreti","section":"§4.1, Figure 13"}],"minor_comments":[{"comment":"The 'original planes 0/191' row is described as the in-build exact reading; the caption should state explicitly that this is not an additional method but the exact ordering used inside trueform before materialisation.","section":"§4.1, Figure 12 caption"},{"comment":"The 'per-edge' column reports the number of correctly partitioned relations, but the column header alone could be misread as the number of failures, especially since the row label 'relations split' precedes it. Add 'correct' to the column header.","section":"Table 3"},{"comment":"The zero-pattern decision tree abbreviates v1,v2,v3 without repeating their definitions in the table. Since the table is meant to be self-contained, consider adding the definitions or a reference to the equations above it.","section":"§2.1, Table 1"},{"comment":"The statement that adjacent-edge correlation 'is local and does not bias the majority' is an assertion, not a demonstrated property. If kept, it should be softened to a stated assumption or supported by the new correlated-corruption experiment recommended above.","section":"Appendix A"},{"comment":"The build/extract timing reports 'the minimum over repeated runs.' A sentence stating the number of repeats per case would help reproducibility; currently only 'best of five' appears for the swarm and browser figures.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong systems contribution with an unusually honest limitations section, but the headline 'exact topology' claim is likely to draw scrutiny from exact-geometry referees. The conditional nature of the majority guarantee should be prominent in the abstract, and the validity oracle should be strengthened or explicitly qualified. The requested changes are local to the framing and the evaluation protocol, not to the core algorithmic idea."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a genuine contribution with unusually honest limitations, but the abstract overstates the guarantee. The core technique is real and the benchmark discipline is among the best I have seen in this area.\n\nWhat is actually new: the radial order around intersection edges is decided by exact signs on the original input planes, without constructing coordinates; disagreeing radial-order observations are aggregated within topological relations by a weighted majority; the arrangement is built per-face with a two-level identity and no global exact structure; and one arrangement build answers arbitrary-arity boolean expressions as per-domain bit tests. The closest prior work (Lévy, EMBER, Cherchi, Böhm and Runge) does not combine these pieces. The paper also ships working code, Python and WebAssembly bindings, and a reproducible benchmark protocol: arrays-in/arrays-out, same machine, same corpus, validity oracle on every output. The 5.5–100× speedups look credible for the tested cases.\n\nThe soft spot is the exactness guarantee itself. Within a single build, the radial order is genuinely exact because it reads original planes. But every cross-face decision—orientation, nesting, relation merges—is resolved by a weighted-majority vote that is correct only when more than half of the weighted observations carry the intended answer. The paper states this in §5.1 and Appendix A, but the abstract still says “we keep it exact” and “produces valid, watertight output.” The validity oracle checks closedness and orientation, not self-intersection, and §5.1 concedes that materialised output can self-intersect. That is not a hidden flaw, but it is an overstatement in the abstract.\n\nThe stress-test concern about correlated flips is fair. Adjacent edges in a relation share endpoints, so one rounded vertex can flip several radial-order readings at once. Under near-coplanar geometry those flips can be systematic, not independent. Appendix A hand-waves this as “local and does not bias the majority” without proving it. The empirical sweeps show the vote tolerates far more construction error than per-edge reading, but they do not establish that real materialisation error is unbiased or that a majority failure is unreachable. The guarantee is statistical, not worst-case.\n\nWho this is for: anyone doing mesh booleans, arrangements, or robust CSG pipelines. It deserves a serious referee. The right referee should ask for a revised abstract, a sharper statement of the majority-vote guarantee, and either a proof or a concrete counterexample for the correlation claim. I would send it to review, and I would expect it to survive with revision.","headline":"Serious, well-engineered contribution; the exactness claim is narrower than the abstract's “we keep it exact,” and the majority-vote guarantee is statistical, but the paper deserves a careful referee.","tokens_in":30327,"tokens_out":3739,"would_cite":true,"duration_ms":40672,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Topological majority votes let mesh booleans stay exact after rounding.","keywords":["constructive solid geometry","mesh boolean operations","topological aggregation","arrangement","exact predicates","materialisation","non-manifold meshes","watertight output"],"falsifier":"Construct a two-sphere arrangement with a ring of near-coplanar faces at the intersection circle, deliberately round the materialised endpoints so that more than half of the edges in one relation flip their radial order, and then check whether the boolean output changes; the paper itself demonstrates the flipping mechanism, so the adversarial test would raise the corruption ratio past 50% and see whether the method still recovers the intended solid.","tokens_in":29540,"feed_emoji":"🔷","tokens_out":2958,"duration_ms":33671,"temperature":0.7,"pith_summary":"This paper argues that the only guarantee a mesh-CSG stage can keep after its output is materialised to floating point is index-based topology, not geometry. It proposes to make that topology exact and to recover disputed geometric decisions through topological aggregation: a weighted-majority vote among the disagreeing observations within each topological unit. If correct, the method produces watertight boolean results without exact constructions, answers booleans of any arity from a single arrangement build, and runs up to two orders of magnitude faster than several prior libraries, including in the browser.","feed_headline":"Majority votes keep mesh booleans exact after rounding","feed_subtitle":"One arrangement build answers any boolean expression, and the topology survives materialised floating-point coordinates.","key_machinery":"Topological aggregation is the central mechanism: each topological unit (a MEL component for orientation, a relation between components for merging, a domain for classification) collects all its geometric observations and resolves disagreement by weighted-majority vote, where weights are the minimum area of incident faces. The supporting machinery is a five-type classification (VV, VE, VF, EE, EF) of pairwise intersections on a bounded integer ladder, a two-level identity scheme (topological identity plus geometric merge) that keeps the arrangement consistent across faces locally, and an implicit reduced graph that carries the same combinatorial content as a Weiler 3-map without explicit hal","core_discovery":"The central claim is that a mesh-CSG pipeline can stay exactly correct by separating what survives materialisation from what does not. Within the build, the radial order around every intersection edge is decided by exact predicates on the original input planes alone, so no constructed coordinate enters the comparison and the result is exact without exact constructions. Where a decision spans faces and individual geometric observations disagree, the intended answer is recovered by a weighted-majority vote over the observations in that topological unit, which the paper models as a maximum-a-posteriori estimate. This makes the arrangement locally computed, globally consistent through identities","pith_inferences":["The majority-vote principle could be applied to other geometric reads that a downstream stage makes from materialised coordinates, such as generalized-winding-number classification or ray parity, not just orientation, merging, and domain classification.","If the build/extract amortisation holds, a natural next target is interactive CSG-tree editing, where each user edit is an extraction from a precomputed arrangement rather than a rebuild.","A concrete stress test suggested by the paper's model: engineer a near-coplanar interface, round the endpoint coordinates to nudge more than half of a relation's edges to the wrong order, and observe whether the output partition flips — this would directly probe the statistical guarantee stated in Appendix A."],"forward_implications":["Any boolean expression over the operands is a per-domain bit test, so a family of CSG queries cost one arrangement build plus cheap extractions.","The radial order around non-manifold edges is exact without exact constructions, because it is computed from original input planes, which removes the need for precision escalation inside the build.","The pipeline keeps partition correctness two to three orders of magnitude longer than per-edge extraction under added construction error, so composed, materialised pipelines can rely on the vote to recover the intended domains.","Reported timings show the method is 5.5× to 100× faster than six prior libraries on a 1000-pair benchmark corpus, and its WebAssembly build stays interactive in the browser on the same inputs."],"fun_headline_variants":["Vote on geometry keeps CSG exact after float rounding","Boolean mesh exactness via majority vote, 100x faster","No exact construction: topological vote for exact CSG","Mesh booleans survive rounding: majority-of-observations wins","Exact CSG without exact coords, built once, any query"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The correctness guarantee is a majority assumption: if materialisation or inconsistent winding corrupts more than half of the observations within a single topological relation, the weighted-majority vote returns the wrong order.","fun_headline_variants_meta":{"raw":{"variants":["Vote on geometry keeps CSG exact after float rounding","Boolean mesh exactness via majority vote, 100x faster","No exact construction: topological vote for exact CSG","Mesh booleans survive rounding: majority-of-observations wins","Exact CSG without exact coords, built once, any query"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000121,"raw_usage":{"total_tokens":937,"prompt_tokens":761,"completion_tokens":176,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":100}},"tokens_in":505,"tokens_out":176,"duration_ms":2877,"temperature":1.0,"reasoning_tokens":100,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:58:00.981515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a two-sphere arrangement with a ring of near-coplanar faces at the intersection circle, deliberately round the materialised endpoints so that more than half of the edges in one relation flip their radial order, and then check whether the boolean output changes; the paper itself demonstrates the flipping mechanism, so the adversarial test would raise the corruption ratio past 50% and see whether the method still recovers the intended solid.","supporting_citations":[],"review_version":1}