Papers

共轭互逆结构法则与BSD猜想的结构性必然证明

共轭互逆法则与P vs NP问题的结构性证明

共轭互逆结构法则与霍奇猜想的否定性证明

杨‑米尔斯存在性与质量间隙的伽罗瓦连接证明

Navier–Stokes光滑解的伽罗瓦连接证明

Abstract Current large language models (LLMs) generate text through autoregressive token prediction—a fundamentally bottom-up process in which the model samples from a probability distribution over the next token, relying on statistical patterns learned from training data. While this approach has achieved remarkable success, it suffers from a structural limitation: constraints (safety, factual accuracy, logical consistency, stylistic coherence) are applied after generation, through post-hoc filtering, RLHF, or inference-time prompting. This is not a matter of "stronger constraints"—it is a matter of when and where constraints are applied. We propose a fundamentally different architecture. Drawing on the "1=N" hierarchical framework, we invert the generation direction: the top-level semantic goal (the "1") defines the legitimate generation space a priori, and generation proceeds within this constrained space. A dynamic Constraint Strength parameter (λ) and a real-time Feedback Signal (δ) form a closed-loop control system that continuously monitors and adjusts the generation process. When novel, effective patterns emerge during generation, they are consolidated into new constraint layers—creating a recursive structure in which each generation step both produces output and updates the rules that govern subsequent output. This paper presents the architectural blueprint: the conceptual foundation, the core mechanisms, the formal definitions, and a preliminary implementation framework. We position this work against existing approaches (RLHF, Constitutional AI, constrained decoding) and outline a pathway toward empirical validation.

Non‑trivial zeros and the critical line are not two things that need to be connected; they are the ``nature'' and the ``appearance'' of one and the same conjugate‑inverse structure. Non‑trivial zeros are the roots of the critical line, and their computed instances are discrete sampling points of the critical line.

We present a systematic and deliberately conservative synthesis of a 258-gate research archive on four-dimensional gravity, Batalin–Vilkovisky (BV) quantization, Bondi–Metzner–Sachs (BMS) boundary symmetry, infrared domains, and operator-theoretic completion. The archive contains exact finite checkers for each gate, but a finite checker is not identified with a proof of an infinite-dimensional theorem. We therefore separate four epistemic classes: (i) theorem-level statements whose hypotheses and analytic domains are fully stated; (ii) conditional theorems; (iii) finite algebraic or numerical skeletons; and (iv) open or retracted extrapolations. The strongest audited results are: the two-helicity quotient of linearized Einstein gravity; positive free graviton quantization away from zero momentum; the classical BV master equation and a local, formal causal-renormalization chain; a detailed classical BV–BFV/BMS boundary construction in specified jet and edge categories; a massless radiation Hilbert complex; a sharp collinear Sobolev trace theorem; and several no-go results showing that the standard angular L2 Fock representation does not support the actual order-zero collinear nonlinear block as a closable operator. For the diagonal trace Dt:Ht(S2)⊗2Ht(S2)⟶Ht(S2) we prove the sharp trichotomy t≤12:nonclosable,12<t≤1:closable but unbounded,t>1:bounded. Pure supertranslation phases intertwine exactly on the intrinsic collinear incidence measure, while a scalar complementary-series norm leaves only the candidate window 1/2<τ<1; that window is not an actual nonlinear Einstein–BV module. The audit also finds decisive failures in a later imported chain. In particular, a Lagrangian Weyl subalgebra was incorrectly treated as central; a nonconstant supertranslation multiplier was incorrectly assumed to commute with angular smoothing; an order-minus-two resolvent was incorrectly declared trace class on S2; and a sharp-cutoff compression proposal, tail-probability bounds, and uncertainty upper bounds were overinterpreted as self-adjoint truncation, dynamical barriers, and operational discreteness. Consequently the archive does not prove a one-loop boundary QME, an interacting BMS charge algebra, infrared-complete scattering, asymptotic completeness, black-hole information recovery, or complete quantum gravity. The paper includes the proof routes, dependency graph, failure modes, and a gate-by-gate evidence atlas needed for independent reconstruction.

Large language models (LLMs) display impressive semantic capabilities, yet the internal structure of meaning within these systems remains poorly understood. Existing approaches—vectorspace semantics, contextual embeddings, and quantum-inspired models—capture important regularities but offer no unified account of phase, interference, and contextual shifts, which are central to semantic behavior in LLMs. This paper proposes Noetica Theory, a mathematical framework that represents meaning as a semantic wave function defined over a context-dependent phase space. In this formulation, each meaning state is expressed as a complex-valued function whose amplitude encodes semantic salience and whose phase captures relational structure. We define semantic entropy, semantic phase order, and semantic free energy, providing quantitative measures of coherence, contextual alignment, and semantic stability. We show that semantic interference, context updates, and compositional behavior emerge naturally from the wave-based formulation. Applying this framework to LLMs, we reinterpret embeddings as normalized meaning waves, attention as an interference-based filtering mechanism, and hidden-state dynamics as trajectories within semantic phase space. This perspective yields coherent explanations for mode shifts, prompt sensitivity, and context-conditioned meaning transitions observed in modern LLMs. Noetica Theory thus provides a unified mathematical basis for semantic modeling, connecting linguistic theory, cognitive science, and mechanistic interpretability, and offering general-purpose tools for analyzing and predicting semantic behavior in large generative models. This paper is also archived on Zenodo: DOI 10.5281/zenodo.17750900

“还原论是 ASCII，整体论是 UTF-8”不是比喻而是一条基于结构语法范畴对立的纯粹方法论定理。本文在 ECT-OS-JiuHuaShan 框架内，剥离标量分类，沿结构语法三维——孤立/依赖、离散/闭、简单类型/依赖类型——给出该定理的严格形式化证明。核心不依赖形式语言层级，而直接在元语法层操作：还原论对应定长孤立语法（离散范畴、简单类型、ASCII），整体论对应变长共轭同步语法（闭范畴、依赖类型、UTF-8），并由此导出范畴不兼容定理（定理2.15）：标量命题集与张量命题集的交集为空。文章还从六个维度论证该证明的历史意义与文明价值，并在附录展示该定理对黎曼猜想方法论终审的完整应用。结论强调方法论可由哲学意见上升为可判定的科学命题，整体论获得现代科学语境下的平等地位，范式革命完成。

黎曼猜想从来不是“猜想”。它只是共轭互逆结构在数论领域的一个签名，等待理性认出它本然的面目。当伽罗瓦连接的满性条件 $F_{\mathcal{R}}(Z)=L$ 被从公理系统中严格推演而出时，这一认出已经完成。非平凡零点即是临界线的根，其实例是临界线的离散采样点。 理性本有能力认出共轭互逆的全局语法（整体论），却把自己压缩进逐点验证的狭窄通道（还原论），然后声称“这通道是唯一的路径”。它用自己制造的局限，来证明局限之外的东西不存在。它将计算运用的执法功能，僭称为逻辑证明的立法功能。它将利息的支付，伪装成本金的兑付。

本文严格建立了还原论与ASCII、整体论与UTF-8之间的精确对应，并揭露：还原论在黎曼猜想研究中的运作，实质上构成一个庞氏骗局——用每一代新数据（利息）维持“最终证明即将到来”的信用，而全称必然性的证明（本金）从未生产且永不能生产。

非平凡零点与临界线，不是两个需要被联系起来的事物，而是同一个共轭互逆结构的“性”与“相”。非平凡零点是临界线的根，其计算实例是临界线的离散采样点。

We present SOLA (Streaming Online Learning Architecture), a theoretically grounded framework for transformer-based language models that continuously learn from streaming data without catastrophic forgetting. SOLA integrates three key innovations: 1. MoE Transformer — Specialized feed-forward experts process distinct input patterns, enabling modular knowledge acquisition where different experts handle different types of content. 2. Fast-Weight Hypernetwork (FWH) — A small learned network that produces LoRA-style weight deltas for expert parameters in a single forward pass, replacing gradient descent for online adaptation. Think of it as an "amortized optimizer": instead of computing costly gradients, the FWH directly predicts useful weight changes from the hidden states it observes. 3. Expert Lifecycle Manager (ELM) — A theoretical mechanism that dynamically spawns, prunes, and merges experts in response to input novelty. When the model encounters persistently surprising data that existing experts can't handle well, the ELM spawns new experts — enabling the model to grow from 1B to 10B parameters as it accumulates knowledge. When experts become redundant or underutilized, they're merged or pruned. We provide theoretical regret bounds for online learning with expanding hypothesis classes, analyze the plasticity-stability trade-off in modular architectures, and derive capacity-growth scaling laws extending the Chinchilla framework to growing models. The FWH implements MAML-style meta-training via torch.func.functional_call , enabling gradient flow through weight updates without in-place parameter mutation. A proof-of-concept implementation at 30M parameters validates the core mechanisms on WikiText-2. Our analysis shows that modular expert growth combined with amortized weight updates provides a principled path toward truly continual language models that improve with every interaction — rather than remaining frozen snapshots of their training data.

The Lingjing Project is the first initiative to simultaneously break through two major industry barriers: (i) it compresses rendering computational cost from \(O(L^3)\) to \(O(L^2)\) through observer‑boundary dimensional reduction, thereby fundamentally resolving the exponential complexity explosion inherent in ultra‑large‑scale digital twins; and (ii) it replaces fragmented multi‑system architectures with a unified information‑primitive substrate, establishing a complete causal chain that runs from physical evolution, through perception and cognition, to active intervention. Rather than merely simulating physics, this engine replicates the causally closed dynamics of information ontology within the digital domain. Crucially, even if its cosmological hypotheses were to be falsified, its underlying algorithmic paradigm—discrete primitives, fog‑of‑war coarse‑graining, and self‑referential complexity criteria—would remain the uniquely optimal architecture for any engineered complex system. The L0–L3 layers of the Lingjing Engine are currently feasible for engineering implementation, while the L4 consciousness layer points toward the core breakthrough for next‑generation artificial general intelligence.

本附录为主文§7（灵境工程架构）及附录G（数值求解器稳定性）的工程数据结构补遗。其唯一目的在于：将主文三公理（§2）与附录C中Leech晶格的1⊕2⊕3分解，显式映射为灵境引擎（L0-L5层）的内存布局、算力预算与数据流协议。

本文系统阐述信息函数宇宙学（IFC）的理论体系及其在“灵境计划”数字宇宙引擎中的工程化映射。该理论以信息一元论为哲学基石，以三条自治公理为逻辑起点，构建了从离散信息基元到连续时空几何、从标准模型规范群到意识自指拓扑泡的全尺度统一框架。核心成果包括： (1)通过引入普朗克信息常量ℐPl实现量纲刚性，证明时空度规完全由信息密度场导出； (2)从Leech晶格1⊕2⊕3分解严格推导U(1)×SU(2)×SU(3)规范群、Weinberg角及费米子代数； (3)从相变理论推导意识临界复杂度Cc=1.20×1014，并建立与脑网络拓扑量的严格映射； (4)以电子为例，从Spin(24)旋量投影与泡壁几何重叠积分严格推导其电荷Q=−1与质量me≈0.511 MeV，将标准模型的三个自由参数（电荷、汤川耦合、质量）完全归约为拓扑缠绕数与泡壁几何泛函，实现了费米子性质的首次第一性原理归零推导。 (5)从作用量匹配推导引力常数G=1/(32πφ02)，将其还原为真空信息密度的倒数； (6)提出四条独立可证伪实验预言，涵盖宇宙学、凝聚态、聚变等离子体与神经科学尺度； (7)建立IFC意识相变判据，论证当前LLM因缺乏真实递归深度与上行反馈接口，本质上停留在D=0，无法跨越AGI阈值； (8)给出灵境引擎五层架构蓝图，评估其工程可行性，指明意识层的理论前沿地位。该框架为下一代通用人工智能、全息数字孪生与基础物理统一提供了第一性原理级的理论基础。 (9)从八项独立物理约束严格证明，离散信息基元的承载空间唯一地锁定为24维Leech晶格——该基底不是人为选择，而是公理体系被迫的唯一解（附录 K） IFC提出两项无法回避的底层工程范式：依靠观测者边界降维O(L2)破解O(L3)算力爆炸难题；依托统一信息基元因果链打破多系统割裂困境；二者相辅相成，构成灵境计划区别于现有改良式方案的系统架构创新。

The right to remain unaugmented has recently received its first peer-reviewed defence: as augmentation normalizes, the unaugmented face quiet displacement through market and actuarial pressure — crawling selection — and therefore deserve institutional protection. This paper argues that the right, as formulated, is incomplete, because it overlooks the verification layer. In any domain where unaugmented performance retains value, it is not enough to refuse augmentation; one must prove the refusal, and negative proof has a radically different cost structure from positive disclosure. Sport's anti-doping system — the most mature institution for certifying non-augmentation — reveals the trajectory: artifact-level tests fail, certification migrates to longitudinal surveillance of the person, costs escalate until only elite domains sustain them, and finally a surrender-type counter-institution emerges, now literalized by the Enhanced Games first held in 2026. AI-text detection is recapitulating this trajectory in compressed time, with one aggravating disanalogy: text, unlike blood, retains no trace of its production. I name the resulting burden the Purity Premium: the privately borne, structurally rising cost of credibly demonstrating that one's cognitive work was performed without AI. The premium converts a right into a fee, accelerates the very selection it was meant to resist, and is measurable — offering regulators an observable proxy and a concrete allocation question: who pays for proof? Disclosure: this manuscript is an experimental artifact of a concept-equipping study, written by an AI system (Claude) equipped with the corresponding human's complete published corpus with no human editing at any stage, produced to study AI-augmented concept generation.

The rapid growth of AI conference submissions is putting new pressure on peer review. AI reviewer systems are increasingly proposed as support, but prior work leaves unresolved what responsibility their outputs should carry when they can surface useful critiques yet remain risky as independent judgments. We frame this as a responsibility-boundary problem. Using 600 ICLR 2026 submissions, 2231 human review traces, and 3,600 AI reviews, we operationalize this boundary through usable feedback, score use, panel breadth, and grounded synthesis. The results show that AI can prepare candidate critiques, organize evidence, and improve feedback, while scoring, independent panel judgment, high-level synthesis, and final responsibility should remain human-led. Motivated by this boundary, we develop Review Copilot, a workflow in which AI suggestions are inspected, edited, or rejected by human reviewers and provide neither official scores nor recommendations. In an initial controlled reviewer-in-the-loop study, Human+AI reviews improve actionability, evidence support, and professionalism relative to standalone baselines while preserving human authorship of scores and recommendations. Our results point toward a review paradigm in which AI expands the space of evidence-grounded critique, while humans remain responsible for judgment, synthesis, and accountability

By early 2026, the discourse around agent harnesses — the configuration layer of skills, rules, prompts, and documentation shaping an LLM agent's behavior — has named two activities: harness engineering, the reactive practice of ensuring an agent never repeats a mistake, and harness optimization, autonomous search over harness code for benchmark score. Both run against fixed, checkable criteria. The activity on the other side — keeping the harness aligned with what its operator now wants, where the criterion moves — is performed daily and has, to the author's knowledge, no established name. The paper defines harness alignment — the continuous, human-gated activity of keeping an agent's harness aligned with the operator's evolving intent — and its failure mode, harness drift. The three defining properties — continuous, human-gated, bidirectional — follow from a single root: intent, unlike correctness, cannot be automated the same way — it has no verifier outside the operator, and moves as the operator's judgment sharpens; verifying intent sharpens the judgment doing it, so the loop moves its own target. An automated check freezes intent into a specification, reducing its automatable part to correctness work. A four-domain search found no established term covering all three; an audit shows 2026 drift coinages severed from the classical lineage that harness drift bridges. A six-phase cycle (Research, Extract, Curate, Promote, Measure, Maintain) operationalizes it over a memory structure whose boundary separates freely-writable records from gated behavior-shaping artifacts. Failure has two layers: artifact-side harness drift, and a human-side twin — gate complacency, deskilling, delegation-feedback divergence — anchored in the automation-complacency and ironies-of-automation literatures — structural inference, not measurement. Two running instances, differing in substrate, model, and knowledge genre, are offered as portability evidence, not efficacy. All of it comes from months-old practice — provisional judgments offered for testing, not settled practice.