{"id":"f587049e-7acd-40c3-9ab3-22a87b7bf7ed","arxiv_id":"2607.03329","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Connection-filtered horizontal learning on an Orlicz fiber bundle bounds OOD predictive variance by the diameter of the identifiable base and eliminates catastrophic forgetting by orthogonal projection of updates.","lead":"The paper defines Statistically Meaningful Geometry (SMG): treat over-parameterized models as infinite-dimensional Orlicz fiber bundles and filter training with an Ehresmann connection so only horizontal statistical directions update. If the geometry is realizable, it claims hard bounds on hallucination variance and non-asymptotic elimination of catastrophic forgetting.","discovery_kind":"paradigm_shift","skeptic_critique":{"model":"grok-4.5","headline":"Theorem 16's hard OOD bound requires a finite-diameter base B that the construction never establishes for real transformers.","rationale":"The reader's weakest assumption correctly flags that no explicit, computable Ehresmann connection ω is supplied for real transformers, so the geometric filter remains formal. That is a genuine engineering gap. The more immediate load-bearing hole for the strongest claim, however, is the finite-diameter hypothesis on B that Theorem 16's proof simply assumes via 'compact embedding.' Without that diameter the OOD variance bound is not hard, and the same object underpins the capacity-collapse and forgetting results. Because the manuscript is pure theory with no constructions, algorithms or experiments, the engineering conclusions remain unsupported; CONDITIONAL is still the right verdict, but the diameter gap should be listed alongside (or ahead of) the missing ω. No change of verdict category is required; the concern sharpens the condition that must be discharged before the central claims can be accepted.","tokens_in":52470,"tokens_out":607,"duration_ms":5593,"concrete_test":"Take a minimal over-parameterized MLP (or 2-layer transformer) with known continuous symmetries (permutation/scaling of hidden units). Explicitly construct the quotient base B of observationally inequivalent input-output maps, equip it with the induced Fisher/L2 metric, and compute or bound diam(B) when the covariate domain is enlarged by a non-empty XOOD of positive measure. If diam(B)=∞ (or grows without bound as the OOD region expands), Theorem 16's hard containment fails for that architecture; if a finite diameter can be exhibited and the filtered flow stays inside it, the claim is locally supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim (Theorem 16) asserts that connection-filtered pre-training strictly upper-bounds OOD predictive variance by the finite diameter of the identifiable quotient base B. The proof (Sec. 8.3) reduces the filtered trajectory to natural-gradient flow on B and then invokes 'standard compact embedding theorems for regular statistical manifolds' so that the path stays inside a compact geodesic ball K whose diameter is finite and independent of ambient weight dimension W. That step is load-bearing: without diam(B)<∞ the claimed hard geometric containment of hallucinations does not follow. The paper never constructs, for a concrete transformer, a base B that is both (i) finite-dimensional and identifiable and (ii) of finite diameter under the induced L2 metric once the domain includes an unbounded OOD set XOOD. The same un-constructed finite-diameter object is reused for the Capacity Collapse Theorem (Thm 12) and the forgetting-elimination argument. The reader's concern about a computable ω is real but secondary; even an oracle connection would leave the diameter claim ungrounded.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Statistically Meaningful Geometry (SMG), lifting over-parameterized models (e.g., transformers) into infinite-dimensional Orlicz statistical manifolds equipped with a general fiber bundle (M, B, π, V, H, ω). An Ehresmann connection 1-form ω filters vertical Structural Internal Directions (SID / gauge noise) so that learning trajectories remain on the horizontal Statistically Verifiable Directions (SVDχ). Under this connection-filtered pre-training the authors claim that out-of-distribution predictive variance is strictly upper-bounded by the finite diameter of the identifiable quotient base B (Theorem 16), thereby geometrically containing generative hallucinations; projecting downstream updates onto the orthogonal complement of the historical horizontal carriage is claimed to yield non-asymptotic elimination of catastrophic forgetting (SMG Sequential Adaptation Flow). Classical MLE, natural gradient, Wilks asymptotics and PAC-Bayesian bounds are recovered on the reduced base, and three nested learning pathways (ambient filtering, leaf integration, macro inference) are shown to be equivalent under integrability.","tokens_in":52735,"tokens_out":1231,"duration_ms":15574,"significance":"If the geometric containment and forgetting-elimination claims hold for realistic architectures, the work would supply a coordinate-free topological foundation for structural reliability in generative AI, resolving the over-parameterization generalization paradox without parameter-counting heuristics. The manuscript supplies an extensive formal apparatus (Lemmas 1–8, Theorems 1–16) that systematically embeds transformers into Pistone–Sempi Orlicz manifolds, defines metric-compatible projections, and derives capacity-collapse and quarantine results; these machine-readable proofs and the clean Two-Fold Inference Paradigm are genuine strengths. The contribution remains purely theoretical: no constructive connection, no numerical verification, and no comparison against existing continual-learning or hallucination-mitigation baselines are provided, so practical impact is still conditional.","major_comments":[{"comment":"§8.3, Theorem 16 (Geometric Containment of Generative Hallucinations): the hard OOD bound reduces the filtered trajectory to natural-gradient flow on B and then invokes “standard compact embedding theorems” so that the path stays inside a compact geodesic ball K of finite diameter KB independent of ambient weight dimension W. The manuscript never constructs, for a concrete transformer, a base B that is simultaneously finite-dimensional, identifiable, and of finite diameter under the induced L2 metric once the evaluation domain includes an unbounded XOOD. Without diam(B)<∞ the claimed geometric containment does not follow; the same ungrounded finite-diameter object is reused for the Capacity Collapse Theorem (Thm 12) and the forgetting argument.","section":"§8.3, Theorem 16"},{"comment":"§§2–4 and 8.3–8.4 (Definitions 9, 20; Theorems 2, 10, 12, 16): all operational claims presuppose a practically usable Ehresmann connection 1-form ω such that the horizontal distribution is Frobenius-integrable (Ω≡0), the orthogonal projection onto the historical horizontal carriage is exact, and the resulting base B has finite diameter. No explicit construction, approximation scheme, or even low-rank surrogate for ω is supplied for transformer weight spaces. Consequently the “total non-asymptotic elimination” of forgetting and the hard hallucination bound remain formal statements conditional on an oracle connection.","section":"§§2–4, 8.3–8.4"},{"comment":"Lemma 4 / Theorem 6 and the Capacity Collapse argument (Thm 12): once the horizontal distribution is defined as the gf-orthogonal complement of ker(dπ) and the connection is defined to project onto that complement, fiber-wise invariance of the likelihood and the collapse of ambient KL divergence to the base KL follow essentially by construction. The manuscript presents these as deep structural theorems; a clearer separation between definitional consequences and non-tautological geometric content would strengthen the load-bearing claims.","section":"Lemma 4, Theorems 6 & 12"}],"minor_comments":[{"comment":"The Two-Fold Inference Paradigm is restated almost verbatim in §§1.3 and 1.7; the four core axioms likewise appear twice with only minor rephrasing. Condensing the introductory material would improve readability.","section":"§§1.3, 1.7"},{"comment":"Figures 1–4 are conceptually helpful but the captions and in-text references contain overlapping descriptive text that is hard to parse; a single clean schematic of the three nested pathways would suffice.","section":"Figures 1–4"},{"comment":"Notation for the horizontal natural gradient (Eq. 87) and the induced base likelihood is introduced late and then used retroactively; a short notation table early in §2 would help.","section":"§5.2, Eq. (87)"},{"comment":"The manuscript cites Amari, Pistone–Sempi and classical PAC-Bayes correctly, yet several recent geometric deep-learning and continual-learning references that address similar gauge/redundancy issues are absent; adding a short related-work paragraph would situate the contribution more clearly.","section":"Introduction / Related work"}],"recommendation":"major_revision","confidential_remarks":"The paper is ambitious and formally dense; its natural home may be a mathematical statistics or information-geometry venue rather than a core cs.LG conference/journal that expects empirical validation. The future-dated arXiv stamp (July 2026) is odd but irrelevant to content. I see no ethical or citation-pattern red flags."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know is that this is a pure-theory attempt to put transformers inside a Pistone–Sempi Orlicz fiber bundle and use an Ehresmann connection to quarantine vertical gauge directions (SID) from horizontal statistical directions (SVDχ). The packaging is careful: the tangent split, the Quarantining Theorem, the Capacity Collapse argument, and the three-way learning hierarchy (ambient filter / leaf / base) are written out with real differential-geometric language rather than slogans.\n\nWhat is actually new is the explicit application of non-principal Ehresmann connections and horizontal-leaf integrability to the two concrete failure modes (hallucination variance and catastrophic forgetting), plus the claim that connection-filtered pre-training bounds OOD predictive variance by diam(B). The classical pieces (likelihood flat on fibers, natural gradient on the quotient, PAC-Bayes after projection) are standard once you accept the split; the synthesis and the named Sequential Adaptation Flow are not routine Amari restatements.\n\nThe soft spots are real but concentrated. Theorem 16’s hard containment reduces the filtered path to natural-gradient flow on B and then invokes compact embedding so that the path lives in a finite-diameter geodesic ball independent of weight dimension W. The paper never constructs, for a concrete transformer, a base B that is both finite-dimensional/identifiable and of finite diameter once XOOD is unbounded. The same unbuilt object underwrites Capacity Collapse and the forgetting-elimination argument. A computable ω is also missing; without it the filter stays formal. There are no experiments, no algorithms, and some duplicated introductory sections. The language (“total non-asymptotic elimination,” “mathematically inevitable”) is stronger than the constructions support.\n\nStill, the skeleton is coherent on its own terms and the infinite-dimensional setup is handled with more care than most ML geometry papers. This is for people who already work with information geometry or gauge ideas in learning theory; they will get a usable vocabulary and a clear list of objects that still need to be built. It is not yet an engineering solution.\n\nI would send it to referees who know both Orlicz manifolds and modern architectures. The central engineering conclusions need heavy revision, but the formal framework is worth the referee time.","headline":"Ambitious fiber-bundle packaging of over-parameterization that formalizes SID/SVDχ and connection filtering, but the hard OOD-diameter and total-forgetting claims rest on unconstructed objects.","tokens_in":53382,"tokens_out":572,"would_cite":false,"duration_ms":7906,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["53C05","62B10","68T07"],"pacs":[],"model":"grok-4.5","headline":"A geometric filter can hard-bound generative hallucination and eliminate catastrophic forgetting in over-parameterized models.","keywords":["Statistically Meaningful Geometry","Ehresmann connection","fiber bundle","over-parameterization","generative hallucination","catastrophic forgetting","Orlicz statistical manifold","PAC-Bayesian bounds"],"falsifier":"Construct an explicit, computable connection for a concrete transformer, train under the filtered flow, and check whether measured out-of-distribution predictive variance stays inside the claimed base diameter while sequential adaptation leaves historical task performance unchanged to machine precision.","tokens_in":53281,"feed_emoji":"📐","tokens_out":636,"duration_ms":4692,"temperature":0.7,"pith_summary":"The paper argues that classical statistics fails for giant over-parameterized models because their parameter spaces are not flat Euclidean containers. They are fiber bundles whose infinite internal degrees of freedom form flat vertical valleys that carry no statistical signal. Those valleys make classical generalization bounds vacuous and let optimization produce generative hallucination and catastrophic forgetting. The proposed Statistically Meaningful Geometry (SMG) framework lifts the model into an infinite-dimensional Orlicz statistical manifold organized as a fiber bundle, then uses an Ehresmann connection as a dynamic filter that discards the vertical gauge noise and keeps learning only on the non-degenerate horizontal directions. Under that filtered pre-training, out-of-distribution predictive variance is proved to be bounded by the finite diameter of the identifiable base manifold, and a sequential adaptation flow that projects new updates off the historical horizontal carriage is proved to eliminate catastrophic forgetting non-asymptotically. The result replaces fine-tuning heuristics with coordinate-free topological constraints.","feed_headline":"Geometry hard-bounds hallucination in giant models","feed_subtitle":"A connection filter keeps learning on the base manifold and kills catastrophic forgetting","key_machinery":"The Ehresmann connection 1-form ω on the SMG fiber bundle (M, B, π, V, H, ω), which splits the tangent space into vertical Structural Internal Directions (SID) and horizontal Statistical Variational Directions (SVDχ) and acts as the geometric filter that quarantines gauge noise from observable learning.","core_discovery":"Under connection-filtered pre-training the model’s out-of-distribution predictive variance is strictly upper-bounded by the finite diameter of the identifiable quotient base manifold, giving a hard geometric containment of generative hallucinations; projecting downstream updates onto the orthogonal complement of the historical horizontal carriage yields total non-asymptotic elimination of catastrophic forgetting.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Connection filter hard-bounds OOD variance by base diameter","SMG projects updates to eliminate catastrophic forgetting","Horizontal distribution confines generative hallucinations","Ehresmann filter strips gauge noise to bound hallucinations","Quotient manifold diameter caps predictive variance hard"],"cache_read_input_tokens":49280,"weakest_assumption_plain":"That a usable Ehresmann connection exists for real transformer architectures so the horizontal directions are integrable, the base is a finite-diameter identifiable manifold, and the orthogonal projection onto the historical horizontal carriage can be computed exactly.","fun_headline_variants_meta":{"raw":{"variants":["Connection filter hard-bounds OOD variance by base diameter","SMG projects updates to eliminate catastrophic forgetting","Horizontal distribution confines generative hallucinations","Ehresmann filter strips gauge noise to bound hallucinations","Quotient manifold diameter caps predictive variance hard"]},"model":"grok-4.5","effort":"low","cost_usd":0.008884,"raw_usage":{"total_tokens":2092,"prompt_tokens":821,"num_sources_used":0,"completion_tokens":70,"cost_in_usd_ticks":88840000,"prompt_tokens_details":{"text_tokens":821,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1201,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":821,"tokens_out":70,"duration_ms":9244,"temperature":1.0,"reasoning_tokens":1201,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T03:14:11.281964+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Construct an explicit, computable connection for a concrete transformer, train under the filtered flow, and check whether measured out-of-distribution predictive variance stays inside the claimed base diameter while sequential adaptation leaves historical task performance unchanged to machine precision.","supporting_citations":[],"review_version":1}