{"id":"3e82cc4e-206a-4ff0-b6a3-2780f4cd0ba6","arxiv_id":"2603.18104","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Composing dimensional types, program hypergraphs, and b-posits yields depth-independent training memory, grade-preserving geometric updates, Bayesian distillation, and certified warm model rotation for domain AI.","lead":"This paper specifies an alternative AI training stack that replaces reverse-mode IEEE-754 autodiff with forward-mode posit/quire arithmetic plus type-level geometric constraints, aiming for ~2× inference memory and grade-preserving updates. It also defines Bayesian distillation from large models and warm rotation for live model swaps with structural certificates.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Prop. 5.1's grade-invariance claim rests on unproven companion invariants and an incomplete dual-number reconstruction argument.","rationale":"The reader's weakest-assumption correctly flags the Bayesian-distillation transfer gap and the dependence on concurrent companion papers, but the more load-bearing soft spot for the strongest claim is the incompleteness of Prop. 5.1 itself: the grade-invariance theorem is the linchpin that turns memory parity and warm-rotation certificates into structural guarantees rather than engineering hopes. The distillation transfer issue is already marked open by the author (Sec. 3.4, 9.1) and is not required for the memory/grade claims. The paper is explicit that implementation is design intent and that present-tense language describes specification, so a CONDITIONAL verdict remains appropriate; the concern does not push to REJECT because the architectural composition is coherent and the failure modes are falsifiable. No change of verdict category is needed, only a sharper focus on the proof obligation that must be discharged before the abstract's 'enables' language is warranted.","tokens_in":22583,"tokens_out":687,"duration_ms":7426,"concrete_test":"Independently re-derive Prop. 5.1 for a single geometric product layer (e.g., 3D PGA bivector weight) without invoking the full PHG saturation engine of [10]: expand the dual-number JVP, form the multi-tangent Gram projection of Sec. 2.4, apply one quire-rounded b-posit update, and check whether any off-grade component appears at machine precision. If a non-zero grade-j component appears for any j \neq k, or if the argument requires an unstated closure property of the dual-number coeffect, the theorem does not hold as stated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim is that PHG grade inference + dual-number forward-mode + quire makes grade preservation a design-time theorem (Prop. 5.1), yielding depth-independent ~2× memory and certified warm rotation. The proof sketch asserts that a grade-k weight has a grade-k tangent, that the random projection v is drawn from the grade-k subspace, and that the reconstructed estimator therefore lies in that subspace. This is only as strong as (i) the coeffect/escape analysis of DTS/DMM [8] and the grade-inference/saturation machinery of PHG [10], both concurrent self-citations with no machine-checked proofs or shipped compiler artifacts in this manuscript, and (ii) the reconstruction step itself: the sketch never shows that the multi-tangent Gram-matrix projection (Sec. 2.4) or the single-tangent estimator remains inside the grade-k subspace after quire-to-posit conversion and the subsequent weight update W' = W - η \nabla W L under finite-precision b-posit arithmetic. If either companion invariant fails to discharge, or if reconstruction/rounding reintroduces off-grade components, the design-time certificate collapses and warm-rotation correctness (Def. 4.1 condition 2) is no longer structural. Bayesian distillation is secondary; the architecture's core theorem is the grade-preservation claim.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Adaptive Domain Models (ADM): a training and deployment architecture that composes three prior results—DTS/DMM forward-mode autodiff with stack-eligible gradients and quire accumulation [8], Program Hypergraph grade inference [10], and b-posit bounded-regime arithmetic [12]—to claim depth-independent training memory of roughly twice the inference footprint, grade-preserving weight updates for Clifford networks, and a unified coeffect signature for STDP neuromorphic learning. It introduces Bayesian distillation (extracting latent priors from general LLMs, then filtering them through DTS/PHG constraints) and warm rotation (atomic, certificate-gated model exchange without service interruption, with signed version records). The central technical claim is Proposition 5.1: under PHG-typed dual-number forward-mode training, grade-k weights retain grade k through updates as a design-time certificate, with sparsity stability and exact rotor equivariance as corollaries.","tokens_in":22926,"tokens_out":1619,"duration_ms":21551,"significance":"If the composition holds in a working compiler and on hardware, the contribution would be substantial for geometric and physics-structured learning: grade corruption under IEEE-754 is a real obstacle to Clifford networks, and a substrate that makes grade preservation a type-level theorem rather than a post-hoc regularizer would change how equivariant geometric models are trained. The memory-parity argument for forward-mode training, if realized, would also make continuous on-device adaptation more plausible on inference-class accelerators. Warm rotation is a clear operational specification that practitioners can evaluate independently of the arithmetic substrate. Credit is due for an honest scope statement (design intent, implementation in progress), for marking the LLM-to-physical-domain transfer gap in §3.4 and §9.1, and for connecting STDP and forward-mode under a shared local-learning coeffect signature. The manuscript does not ship machine-checked proofs, trained models, or measured memory/accuracy results; significance therefore remains conditional on the companion frameworks and on empirical validation that is not yet present.","major_comments":[{"comment":"§5.1, Proposition 5.1: The grade-invariance claim is load-bearing for the paper’s strongest result (certified warm rotation via PHG certificates, Def. 4.1 condition 2). The proof sketch correctly notes dual-number grade matching and grade-restricted projections, but it does not close the argument under the multi-tangent Gram projection of §2.4, quire-to-posit conversion, or the finite-precision update W′=W−η∇_W L in b-posit. Off-grade components reintroduced by rounding or reconstruction would falsify the design-time certificate. Either complete the reconstruction/rounding argument in this manuscript or state Prop. 5.1 as conditional on fully discharged invariants from [8,10] with an explicit residual risk.","section":"§5.1, Proposition 5.1"},{"comment":"Dependence on concurrent self-citations [8] and [10]: Stack-eligible O(1)/O(k) coeffects, quire lifetime analysis, PHG grade inference, and SMT-LIB2 discharge of rotor norms are imported as established, yet this manuscript provides neither machine-checked artifacts nor independent restatements sufficient to audit those claims. For a journal contribution whose memory-parity (~2× inference) and grade-preservation theorems rest on those results, the paper must either (a) include self-contained formal statements and proof obligations that a reader can check without the companions, or (b) defer the theorem-level claims until the companions are available with verifiable artifacts, and reframe the present work as an architectural specification.","section":"§1.2, §2.3–2.4, §5.1"},{"comment":"Absence of empirical or systems evidence: The abstract and §4 claim depth-independent training memory, productive spare-TOPS training (Fig. 2), stable 85–95% Cayley sparsity after training, and continuous adaptation. There are no trained models, memory footprints, accuracy curves, KL-trigger traces, or certificate-diff examples. For an architecture paper whose central selling points are quantitative (2× memory, grade stability, warm rotation without interruption), at least one end-to-end case study—even a small PGA network or a simulated rotation on a fixed hardware budget—is needed to show that the design intent is realizable and that Prop. 5.1 survives implementation.","section":"Abstract; §4.2; Fig. 2; §5.1 Corollary 5.1"},{"comment":"§3.4 Bayesian distillation: The paper claims the mechanism “resolves the data-scarcity bootstrapping problem,” yet §3.4 and §9.1 correctly note that van Steenkiste & Linzen [26] evaluate sequential recommendation, not continuous multi-physics grade-structured weight spaces, and that transfer is open. The type-to-support construction (W_adm = W_dim ∩ W_grade, projector Π_adm) is a useful formalization, but calling the problem resolved overstates the evidence. Soften the claim to a proposed initialization path contingent on the open transfer questions, and separate what is formal (filter support) from what is empirical (calibration of the distilled prior).","section":"§3.4; §9.1; Contribution 3"}],"minor_comments":[{"comment":"Present-tense system behavior in §4 (“the active model actor processes…”) is later qualified as design intent; a consistent “specification / intended behavior” framing would reduce reader confusion.","section":"§4 opening; Def. 4.1"},{"comment":"Typo in Prop. 5.1: “as a a design-time certificate.”","section":"§5.1"},{"comment":"Fig. 1 and Fig. 2 are conceptual only; adding a small worked numerical example (e.g., one layer’s grade-typed dual update) would help readers who are not already steeped in the companion papers.","section":"Fig. 1; Fig. 2"},{"comment":"Free parameters ε_domain, k (tangents), r (Hessian rank), and b-posit eS/rS are introduced without recommended defaults or sensitivity discussion; a short table of suggested ranges would aid reproducibility of any future experiments.","section":"§2.4; §3.2; §4.2"},{"comment":"Software Availability points to repositories “under active development”; stating which claims are backed by public code versus specification-only would set expectations clearly.","section":"Software Availability"}],"recommendation":"major_revision","confidential_remarks":"Fit is borderline for a results-oriented AI journal: the manuscript is a design composition paper that leans heavily on two concurrent, same-author works not yet independently available. If the venue expects empirical ML systems results, major revision may still fail without a minimal prototype. If the venue accepts formal systems architecture with deferred implementation, the revision path above is appropriate. Citation pattern is dominated by the author’s own sequence plus a small set of geometric-algebra and posit references; that is understandable for a framework paper but will draw scrutiny. I do not see internal contradiction so much as incomplete discharge of the central theorem."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline: this is a systems-architecture design paper, not a results paper. It composes the author’s concurrent DTS/DMM and PHG work with b-posit and multi-tangent forward gradients into a single story for geometric and neuromorphic domain models. The named mechanisms—Bayesian distillation (LLM latent prior filtered by type/grade constraints) and warm rotation (certified atomic live weight exchange with PHG diffs and signed version records)—are the actual packaging contributions. They are not already sitting in the geometric-algebra or neuromorphic literature under those names.\n\nWhat it does well: the problem framing is clear and fair. IEEE-754 grade corruption, reverse-mode tape cost, and the self-reinforcing contamination of Clifford weights are real issues; the paper states them without melodrama. The memory-parity argument (~2× inference footprint, depth-independent) follows standard forward-mode coeffect reasoning if the cited stack-eligibility and quire results hold. Prop. 5.1’s sketch is directionally right for grade-closed dual numbers. The STDP/forward-mode coeffect unification and the actor-model placement between MoE and agentic frameworks are useful conceptual moves. Citations to Baydin, Flügel, Ruhe, Zhdanov, Gustafson, and van Steenkiste are appropriate; self-citation of [8,10] is load-bearing but not decorative.\n\nSoft spots, in proportion: there are no experiments, no trained models, no measured memory or accuracy numbers, and the text repeatedly says implementation is in progress and present-tense behavior is design intent. Prop. 5.1 is a proof sketch that assumes the companion grade-inference and coeffect machinery discharge, and it does not fully close the reconstruction/rounding step under finite-precision b-posit after the Gram projection or the weight update. That is a real gap for a claim sold as a design-time theorem. Bayesian distillation’s transfer assumption (sequential-recommendation LLM structure → continuous multi-physics grade-structured weights) is outside the cited empirical regime; the paper itself marks this as open. Free parameters (ε_domain, k, posit eS/rS, Hessian rank) are acknowledged but not fixed by data.\n\nWho it is for: people building geometric/equivariant or neuromorphic training stacks, or thinking about certified live model updates. Not for someone who needs a trained baseline tomorrow. I would send it to peer review as a design/theory piece that needs sharper theorem statements and an honest “implementation status” section; I would not desk-reject it. Worth engaging if you care about the substrate problem; not yet citable as demonstrated fact.","headline":"A coherent design paper that packages grade-preserving forward-mode training, warm rotation, and Bayesian distillation—but the load-bearing theorem is a sketch on unshipped companion invariants, with no experiments.","tokens_in":23558,"tokens_out":635,"would_cite":false,"duration_ms":6727,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Training memory, grade corruption, and optimizer bloat are substrate artifacts of reverse-mode IEEE-754, not of learning itself; a type-level geometric substrate can train at ~2× inference memory with certified structure.","keywords":["adaptive domain models","forward-mode autodiff","geometric algebra","grade preservation","program hypergraph","warm rotation","Bayesian distillation","neuromorphic STDP"],"falsifier":"Train a Clifford (or hybrid neuromorphic) network end-to-end under the claimed substrate and check two concrete outcomes: (1) every weight remains pure grade after many steps (Cayley sparsity and rotor norms identical to initialization), and (2) peak training memory stays within a small constant of twice the inference footprint independent of depth; failure of either measurement falsifies the central claim.","tokens_in":23372,"feed_emoji":"⚙️","tokens_out":753,"duration_ms":6654,"temperature":0.7,"pith_summary":"The paper argues that the familiar costs of deep training—activation-tape memory that scales with depth, optimizer state that doubles parameters, and the silent destruction of geometric algebra structure—are consequences of reverse-mode autodiff over IEEE-754 arithmetic, not of gradient-based learning. It composes three prior results (dimensional types with stack-eligible gradients and exact quire accumulation, a program hypergraph that treats grade as a type invariant, and bounded-regime posit arithmetic) into a single training regime. Under that regime, a weight of declared grade keeps a same-grade gradient by construction, so trained Clifford networks retain the sparsity and equivariance they were designed for; training memory stays roughly twice the inference footprint regardless of depth. The same coeffect signature covers both loss-function networks and spike-timing plasticity, so geometric and neuromorphic models share one verification and deployment stack. Two operational pieces complete the picture: Bayesian distillation, which filters a general model’s latent prior through the type system to seed domain models when data are scarce, and warm rotation, which swaps a certified updated model into a live inference path without interrupting service. The intended outcome is smaller, continuously adaptive domain systems that are verifiably consistent with the physics of their domains and can be bootstrapped from existing large models.","feed_headline":"Train geometric AI at twice inference memory, grades intact","feed_subtitle":"Type-level grade rules and exact quire gradients replace reverse-mode IEEE-754 overhead","key_machinery":"Adaptive Domain Model (ADM) training substrate: PHG grade inference plus dual-number forward gradients with exact quire accumulation, so that a grade-k weight has a grade-k gradient by the chain-rule closure of grade-preserving operations and any other grade is rejected at elaboration time.","core_discovery":"Composing program-hypergraph grade inference, dual-number forward-mode autodiff, and quire accumulation turns grade preservation through training into a design-time theorem, bounds training memory to about twice the inference footprint independent of depth, and thereby supports certified warm rotation and Bayesian distillation for domain-specific geometric and neuromorphic models.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Grade-preserving training at ~2x inference memory","Exact quire gradients keep geometric grades intact","Bayesian distillation seeds domain models from priors","Warm rotation swaps models with PHG certificates","Depth-independent memory for geometric neuromorphic AI"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That latent Bayesian structure extractable from general language models will, after dimensional and grade filtering, become a well-calibrated prior on continuous multi-physics weight spaces—an assumption the paper itself places outside the cited empirical regime and lists as open work—while the whole architecture also assumes the concurrent type-system and hypergraph results deliver their claimed coeffect and grade invariants in a working compiler.","fun_headline_variants_meta":{"raw":{"variants":["Grade-preserving training at ~2x inference memory","Exact quire gradients keep geometric grades intact","Bayesian distillation seeds domain models from priors","Warm rotation swaps models with PHG certificates","Depth-independent memory for geometric neuromorphic AI"]},"model":"grok-4.5","effort":"low","cost_usd":0.00304,"raw_usage":{"total_tokens":1138,"prompt_tokens":851,"num_sources_used":0,"completion_tokens":52,"cost_in_usd_ticks":30400000,"prompt_tokens_details":{"text_tokens":851,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":235,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":851,"tokens_out":52,"duration_ms":2720,"temperature":1.0,"reasoning_tokens":235,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T23:02:45.192338+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train a Clifford (or hybrid neuromorphic) network end-to-end under the claimed substrate and check two concrete outcomes: (1) every weight remains pure grade after many steps (Cayley sparsity and rotor norms identical to initialization), and (2) peak training memory stays within a small constant of twice the inference footprint independent of depth; failure of either measurement falsifies the central claim.","supporting_citations":[],"review_version":2}