math-skill 2.0.1 → 3.0.0

package/lenses/game.en.md

@@ -0,0 +1,50 @@
+# 🎯 Game Lens
+
+> Your optimal choice depends on others' choices — thinking must be not only deep but interactive.
+
+## What Perspective It Offers
+
+This is a "player's" perspective — in environments where multiple decision-makers mutually influence one another, each participant's optimal action depends on others' actions, and others' optimal actions in turn depend on yours. It demands recursive depth of reasoning: "I think that you think that I think..." The core is not optimizing one's own payoff, but finding a stable state from which no one is willing to unilaterally deviate — the Nash equilibrium.
+
+## What Problems It Is Suited to Diagnose
+
+- Multiple decision-makers whose strategies affect one another's payoffs
+- Designing incentive mechanisms that channel self-interested behavior toward desired social outcomes
+- Analyzing optimal strategies and equilibrium predictions in competitive or adversarial scenarios
+- Anticipating others' actions and adjusting one's own decisions accordingly
+
+## What Problems It Is Not Suited For
+
+- Single-agent decision problems with no interaction with others — optimization thinking is more appropriate
+- Purely cooperative problems where all participants share identical objectives — strategic analysis is unnecessary
+- Deterministic problems where outcomes are uniquely determined by one's own actions — no others' responses are involved
+- Outcomes determined purely by luck, with no strategic choices available to any party
+
+## Which Knowledge Domains It Routes To
+
+- Game theory (Nash equilibrium / dominant strategies / mixed strategies): the core framework for analyzing strategic interaction
+- Mechanism design: redesigning the rules when equilibria are undesirable, making cooperation the self-interested choice
+- Optimization theory (minimax / linear programming): computational tools for zero-sum games and equilibrium computation
+- Probability theory (Bayesian games / expected payoffs): handling asymmetric information and type uncertainty
+
+## What AI Designs It May Inspire
+
+- Game-type classifier: zero-sum / non-zero-sum / sequential / cooperative → automatic matching of solution methods
+- Equilibrium solver and stability checker: Nash equilibrium + trembling-hand perfection + evolutionary stability
+- Mechanism design engine: when equilibria are undesirable, automatically generating payoff-structure / rule-modification recommendations
+
+## Reasoning Protocol
+
+1. **Identify players and strategy sets**: Who is making decisions? What can each choose? Omitting a key player is the most common error
+2. **Determine the game type**: Zero-sum / general-sum? Simultaneous / sequential? Complete / incomplete information? The type determines the method
+3. **Construct the payoff structure**: Write out the payoff matrix or payoff functions; verify completeness and accuracy
+4. **Solve for equilibrium**: Dominant strategies → pure-strategy Nash → mixed strategies → subgame perfection; rank multiple equilibria by Pareto dominance
+5. **Test stability and design mechanisms**: Is the equilibrium robust? If the equilibrium is undesirable, redesign the rules so that good outcomes become the self-interested choice
+
+## Acceptance Criteria
+
+- All players have been identified; strategy sets and payoff functions are explicitly defined
+- The game type has been determined and the corresponding solution method selected
+- Equilibria have been computed (pure / mixed / subgame-perfect); multiple equilibria are annotated and ranked
+- Stability checks have been performed (trembling-hand / evolutionary stability / deviation incentives)
+- If equilibria are undesirable, mechanism design recommendations have been provided (rule changes / payoff modifications / reputation mechanisms)

package/lenses/game.md

@@ -0,0 +1,50 @@
+# 🎯 博弈透镜 / Game Lens
+
+> 你的最优选择取决于他人的选择——思考不仅要深入，还要互动。
+
+## 它是什么视角
+
+这是一种"对弈者"的视角——在多个决策者相互影响的环境中，每个参与者的最优行动取决于他人的行动，而他人的最优行动又反过来取决于你。它要求思维具有递归深度："我认为你认为我认为……"。核心不是优化自己的收益，而是找到无人愿意单方面偏离的稳定状态——纳什均衡。
+
+## 适合诊断什么问题
+
+- 多个决策者相互影响，各自策略影响他人收益
+- 需要设计激励机制，使自利行为导向期望的社会结果
+- 分析竞争/对抗场景中的最优策略与均衡预测
+- 预判他人行动并据此调整自己的决策
+
+## 不适合什么问题
+
+- 单人决策问题，与他人无互动——优化思想更适用
+- 纯合作问题，参与者目标完全一致——无需战略分析
+- 确定性问题，结果由自身行动唯一决定——不涉及他人反应
+- 纯粹运气决定的结果，各方无策略选择
+
+## 会路由到哪些知识域
+
+- 博弈论（纳什均衡/占优策略/混合策略）：分析战略互动的核心框架
+- 机制设计：当均衡不理想时重新设计规则，让合作成为自利选择
+- 优化理论（minimax/线性规划）：零和博弈与均衡求解的计算工具
+- 概率论（贝叶斯博弈/期望支付）：处理信息不对称与类型不确定性
+
+## 可能产生哪些 AI 设计
+
+- 博弈类型分类器：零和/非零和/序列/合作→自动匹配求解方法
+- 均衡求解与稳定性检验：纳什均衡 + 颤抖手完美性 + 演化稳定性
+- 机制设计引擎：均衡不理想时，自动生成支付结构/规则修改建议
+
+## 推理协议
+
+1. **识别参与者与策略集**：谁在决策？每人可选什么？遗漏关键参与者是最常见错误
+2. **判定博弈类型**：零和/一般和？同时/序列？完全/不完全信息？类型决定方法
+3. **构建支付结构**：写出支付矩阵或支付函数，验证完备性与准确性
+4. **求解均衡**：占优策略→纯策略纳什→混合策略→子博弈完美，多重均衡做帕累托排序
+5. **检验稳定性与设计机制**：均衡是否鲁棒？若均衡不理想，重新设计规则让好结果成为自利选择
+
+## 验收标准
+
+- 所有参与者已识别，策略集与支付函数已显式定义
+- 博弈类型已判定，对应求解方法已选择
+- 均衡已计算（纯/混合/子博弈完美），多重均衡已标注并排序
+- 稳定性检验已执行（颤抖手/演化稳定/偏离激励）
+- 若均衡不理想，机制设计建议已给出（改规则/改支付/引入声誉）

package/lenses/geometric.en.md

@@ -0,0 +1,53 @@
+# 🌉 Geometric Lens
+
+> Metric, curvature, and spatial structure on manifolds — seeing problems through the eyes of differential geometry, where distance, bending, and optimal paths reveal the intrinsic shape of data
+
+## What This Perspective Is
+
+This is a "surveyor's" perspective — endowing the problem space with a metric tensor (defining distance and angle), characterizing spatial bending through curvature, and finding optimal paths via geodesics. The core conviction: many seemingly complex nonlinear problems become natural once the correct manifold structure is recognized — the geometry of parameter space determines the optimization landscape, and the intrinsic dimensionality of data sets the ultimate compression limit.
+
+## Problems It Diagnoses Well
+
+- The geometry of parameter space — is a Euclidean metric adequate, or is a Riemannian metric needed?
+- The intrinsic dimensionality of data is far below the embedding dimension — does the manifold hypothesis hold?
+- Distance between two distributions or models — the difference between Fisher metric and Euclidean distance
+- Optimization landscape analysis — high-curvature directions converge fast, low-curvature directions converge slowly
+- The importance of preserving intrinsic geometry in representation learning
+
+## Problems It Doesn't Fit
+
+- Problems with no natural geometric structure — forcing a metric only introduces spurious structure
+- Purely algebraic or combinatorial problems — no concept of distance, angle, or bending is involved
+- Manifold dimension close to embedding dimension — the manifold hypothesis offers no compression benefit
+
+## Knowledge Domains It Routes To
+
+- **differential-geometry/manifold**: Manifold definitions, atlases, coordinate transformations — the stage for geometric reasoning
+- **differential-geometry/metric-tensor**: Metric tensors defining distance and inner product; Fisher information matrix as the natural metric
+- **differential-geometry/curvature**: Curvature tensors and sectional curvature — precise measures of spatial bending
+- **differential-geometry/geodesic**: Geodesic equations and exponential maps — shortest paths on manifolds
+- **optimization/riemannian-optimization**: Riemannian gradient descent and retraction operators — optimization algorithms on manifolds
+- **information-geometry/natural-gradient**: Natural gradient as steepest descent direction under the Fisher metric
+
+## AI Designs It May Inspire
+
+- **Riemannian Optimizer**: Perform natural gradient descent on parameter spaces with manifold constraints
+- **Geometry-Aware Attention**: Replace Euclidean distance with geodesic distance when computing attention weights
+- **Manifold Regularization**: Add curvature penalties to the loss function to maintain geometric smoothness of the representation space
+- **Intrinsic Dimension Estimator**: Estimate the intrinsic dimension of the data manifold online to guide bottleneck layer width selection
+
+## Reasoning Protocol
+
+1. **Identify the manifold structure**: What is the natural parameter space? What type of manifold is it (sphere, Grassmannian, SPD matrix space)?
+2. **Choose a metric**: Is the Euclidean metric sufficient, or is a Fisher information metric or other Riemannian metric required?
+3. **Analyze curvature**: What is the curvature of the space? Positive (spherical) vs. negative (hyperbolic) vs. flat — what does this imply for optimization and representation?
+4. **Compute geodesics**: What is the optimal path between two points? Can the exponential and logarithmic maps be computed efficiently?
+5. **Design geometry-consistent algorithms**: Ensure optimization steps respect the manifold structure (retraction to manifold, Riemannian gradient)
+
+## Acceptance Criteria
+
+- The manifold type and coordinate parameterization are clearly defined
+- The metric tensor has been chosen with theoretical or empirical justification
+- Curvature properties have been analyzed and their implications for optimization and representation are annotated
+- Geodesic distance and exponential maps have a computationally feasible scheme
+- Algorithm design respects manifold structure — intrinsic vs. extrinsic quantities are distinguished

package/lenses/geometric.md

@@ -0,0 +1,53 @@
+# 🌉 几何透镜
+
+> 度量/曲率/流形上的空间结构——用微分几何的眼光看问题，距离、弯曲与最优路径揭示数据的内在形状
+
+## 它是什么视角
+
+这是一种"测量者"的视角——为问题空间赋予度量张量（定义距离与角度），用曲率刻画空间的弯曲程度，用测地线寻找最优路径。核心信念：许多看似复杂的非线性问题，在正确的流形结构下会变得自然——参数空间的几何决定了优化的地形，数据的内在维度决定了可用的压缩极限。
+
+## 适合诊断什么问题
+
+- 参数空间的几何结构——欧氏度量是否合适？是否需要黎曼度量？
+- 数据的内在维度远低于嵌入维度——流形假设是否成立？
+- 两个分布/模型之间的距离——Fisher 度量 vs 欧氏距离的差异
+- 优化地形分析——曲率大的方向收敛快，曲率小的方向收敛慢
+- 表示学习中保持内蕴几何不变的重要性
+
+## 不适合什么问题
+
+- 问题本身没有自然的几何结构——强行赋予度量只会引入伪结构
+- 纯代数/组合问题——不涉及距离、角度、弯曲的概念
+- 流形维度接近嵌入维度——流形假设无压缩收益
+
+## 会路由到哪些知识域
+
+- **differential-geometry/manifold**：流形定义、图册、坐标变换——几何推理的舞台
+- **differential-geometry/metric-tensor**：度量张量定义距离与内积，Fisher 信息矩阵作为自然度量
+- **differential-geometry/curvature**：曲率张量、截面曲率——空间弯曲的精确度量
+- **differential-geometry/geodesic**：测地线方程、指数映射——流形上的最短路径
+- **optimization/riemannian-optimization**：黎曼梯度下降、收缩算子——流形上的优化算法
+- **information-geometry/natural-gradient**：自然梯度 = Fisher 度量下的最速下降方向
+
+## 可能产生哪些 AI 设计
+
+- **黎曼优化器**：在流形约束的参数空间上做自然梯度下降
+- **几何感知注意力**：用测地距离替代欧氏距离计算注意力权重
+- **流形正则化**：损失函数中加入曲率惩罚，保持表示空间的几何光滑性
+- **内蕴维度估计器**：在线估计数据流形的本征维度，指导瓶颈层宽度选择
+
+## 推理协议
+
+1. **识别流形结构**：问题的自然参数空间是什么？它是哪种流形（球面、Grassmannian、SPD 矩阵空间）？
+2. **选择度量**：欧氏度量是否充分？是否需要 Fisher 信息度量或其他黎曼度量？
+3. **分析曲率**：空间的曲率如何？正曲率（球面）vs 负曲率（双曲）vs 平坦——对优化和表示意味着什么？
+4. **计算测地线**：两点之间的最优路径是什么？指数映射和对数映射是否可高效计算？
+5. **设计几何一致的算法**：确保优化步骤尊重流形结构（收缩到流形、黎曼梯度）
+
+## 验收标准
+
+- 流形类型与坐标参数化已明确定义
+- 度量张量已选择且有理论或经验依据
+- 曲率性质已分析，其对优化/表示的影响已标注
+- 测地距离/指数映射有可计算方案
+- 算法设计尊重流形结构——内蕴量 vs 外蕴量已区分

package/lenses/local-to-global.en.md

@@ -0,0 +1,52 @@
+# 📈 Local-to-Global Lens
+
+> Patching local properties into global, sheaf cohomology obstructions — can local solutions be seamlessly assembled into a global solution? Where do the obstructions lie?
+
+## What This Perspective Is
+
+This is a "patcher's" perspective — starting from local patches that cover a space, asking: "Can properties that hold in each local region be consistently assembled into a globally valid property?" The core conviction: the difficulty of many global problems lies not in local solving but in the compatibility between local solutions. The partition of unity theorem guarantees that smooth functions can always be patched from local to global, but sheaf cohomology precisely characterizes when such patching fails — a nonzero cohomology group signals the existence of a global obstruction.
+
+## Problems It Diagnoses Well
+
+- Whether locally defined quantities (gradients, features, representations) can be consistently assembled into a global quantity
+- When analytic continuation is unique and when multivaluedness arises (applicability of the monodromy theorem)
+- Whether local consensus in a distributed system can guarantee global consistency
+- Whether local trivializations of a fiber bundle can be patched into a global trivialization — what is the topological obstruction?
+- Detecting global inconsistencies across overlapping covers via Cech cohomology
+
+## Problems It Doesn't Fit
+
+- Problems that are inherently global and indivisible — no natural local cover exists
+- Concerns only single-point properties rather than the local-to-global transition
+- Purely algebraic problems — no concept of spatial covering or patching is involved
+
+## Knowledge Domains It Routes To
+
+- **topology/persistent-homology**: Persistent homology — a bridge from local neighborhoods to global topological features
+- **topology/fundamental-group**: Fundamental group as a global invariant of path connectivity, detecting "holes" invisible locally
+- **topology/euler-characteristic**: Euler characteristic — linking local combinatorial information to global topology
+- **differential-geometry/connection**: Connections defining parallel transport — consistency conditions from local tangent spaces to global
+- **differential-geometry/manifold**: Atlases and coordinate transformations — the manifold definition itself is the paradigm of local-to-global
+
+## AI Designs It May Inspire
+
+- **Sheaf Consistency Loss**: Penalize incompatibility between representations in adjacent local regions, driving globally consistent representation learning
+- **Local-Global Verifier**: Solve independently on multiple patches, then check global consistency via the Cech condition
+- **Cohomological Obstruction Detector**: Automatically identify which local solution combinations have topological obstructions requiring global correction
+- **Partition-and-Stitch Framework**: Decompose a global problem into a weighted superposition of local subproblems using the partition of unity idea
+
+## Reasoning Protocol
+
+1. **Choose a local cover**: What family of open sets covers the problem space? What are the granularity and overlap of the cover?
+2. **Solve local problems**: Solve independently on each covering set, recording the form and scope of each solution
+3. **Check compatibility conditions**: Are the local solutions consistent on overlap regions? Do the transition functions satisfy the cocycle condition?
+4. **Identify cohomological obstructions**: If compatibility fails, compute the Cech cohomology group — a nonzero group indicates an irremovable global obstruction
+5. **Decide on a strategy**: Is the obstruction removable (adjust local solutions) or essential (a global method is required, or multivaluedness must be accepted)?
+
+## Acceptance Criteria
+
+- The local cover is clearly defined with stated granularity and overlap
+- Each local solution is independently provided with its scope annotated
+- Compatibility conditions on overlap regions have been checked and results recorded
+- If obstructions exist, the cohomology group has been computed or estimated and the obstruction type classified
+- A final strategy is determined — patching succeeded, global correction needed, or obstruction is irremovable

package/lenses/local-to-global.md

@@ -0,0 +1,52 @@
+# 📈 局部到全局透镜
+
+> 局部性质拼接为全局、层上同调障碍——局部解能否无缝组装成全局解？障碍在哪里？
+
+## 它是什么视角
+
+这是一种"拼接者"的视角——从覆盖空间的局部 patches 出发，追问"每个局部区域内成立的性质，能否一致地拼接为全局成立的性质？"核心信念：许多全局问题的难度不在于局部求解，而在于局部解之间的相容性。单位分解定理保证了光滑函数总可局部拼接为全局，但层上同调精确刻画了什么时候这种拼接会失败——上同调群非零 = 全局障碍存在。
+
+## 适合诊断什么问题
+
+- 局部定义的量（梯度、特征、表示）能否一致地组装为全局量？
+- 解析延拓何时唯一、何时出现多值性（单值化定理的适用性）
+- 分布式系统中局部共识能否保证全局一致性？
+- 纤维丛的局部平凡化能否拼接为全局平凡化——拓扑障碍是什么？
+- Cech 上同调检测覆盖之间的全局不一致性
+
+## 不适合什么问题
+
+- 问题本身就是全局的且不可分解——没有自然的局部覆盖
+- 只关心单点性质而非局部到全局的过渡
+- 纯代数问题——不涉及空间覆盖与拼接的概念
+
+## 会路由到哪些知识域
+
+- **topology/persistent-homology**：持续同调——从局部邻域到全局拓扑特征的桥梁
+- **topology/fundamental-group**：基本群作为路径连通性的全局不变量，检测局部无法捕获的"洞"
+- **topology/euler-characteristic**：Euler 示性数——局部组合信息与全局拓扑的关联
+- **differential-geometry/connection**：联络定义平行移动——局部切空间到全局的一致性条件
+- **differential-geometry/manifold**：图册与坐标变换——流形定义本身即局部到全局的范式
+
+## 可能产生哪些 AI 设计
+
+- **层一致性损失**：惩罚相邻局部区域表示之间的不相容，驱动全局一致表示学习
+- **局部-全局验证器**：在多 patch 上独立求解后用 Cech 条件检验全局一致性
+- **上同调障碍检测器**：自动识别哪些局部解组合存在拓扑障碍，需要全局修正
+- **分区拼接框架**：用单位分解思想将全局问题分解为加权局部子问题的叠加
+
+## 推理协议
+
+1. **选择局部覆盖**：用什么开集族覆盖问题空间？覆盖的粒度和重叠度如何？
+2. **求解局部问题**：在每个覆盖集上独立求解，记录解的形式与适用范围
+3. **检查相容性条件**：在重叠区域，局部解是否一致？转移函数是否满足余环条件？
+4. **识别上同调障碍**：若相容性失败，计算 Cech 上同调群——非零群 = 不可消除的全局障碍
+5. **决定策略**：障碍可消除（调整局部解）还是本质性的（需要全局方法或接受多值性）？
+
+## 验收标准
+
+- 局部覆盖已明确定义，覆盖粒度与重叠度已说明
+- 每个局部解已独立给出并标注适用范围
+- 重叠区域的相容性条件已检查，结果已记录
+- 若存在障碍，上同调群已计算或估计，障碍类型已分类
+- 最终策略已确定——拼接成功/需要全局修正/障碍不可消除

package/lenses/perturbation.en.md

@@ -0,0 +1,51 @@
+# 🧠 Perturbation Lens
+
+> Propagation of small perturbations, stability, and robustness — a system's sensitivity to tiny changes determines its reliability
+
+## What This Perspective Is
+
+This is a "stress-tester's" perspective — for any mathematical object or system design, ask: "If the input, parameters, or structure undergo a small change, how much does the output shift? Is the shift controllable?" The core conviction: a well-posed system's response to perturbation should be bounded and predictable; if a design collapses under infinitesimal perturbation, it is destined to fail in practice. The condition number is the precise measure of perturbation amplification.
+
+## Problems It Diagnoses Well
+
+- Robustness analysis of models against input noise or adversarial perturbations
+- Small parameter changes causing solution jumps in optimization — is the problem ill-conditioned?
+- Stability of eigenvalues and eigenvectors under matrix perturbation (Davis-Kahan theorem)
+- Identifying regular vs. singular perturbations in asymptotic expansions
+- Error propagation and condition number analysis in numerical algorithms
+
+## Problems It Doesn't Fit
+
+- Problems where the perturbation itself is the object of study (e.g., butterfly effect in chaotic systems is a feature, not a defect)
+- Large-deformation or global structural analysis — perturbation theory only concerns local neighborhoods
+- Discrete combinatorial problems — perturbation theory relies on continuity assumptions
+
+## Knowledge Domains It Routes To
+
+- **matrix-analysis/matrix-perturbation**: Weyl's inequality, Davis-Kahan theorem, perturbation bounds for eigenvalues and singular values
+- **optimization/constrained-optimization**: Stability of KKT conditions, sensitivity of optimal solutions to perturbation
+- **probability/concentration-inequality**: Concentration phenomena under random perturbation, sub-Gaussian tail bounds
+- **differential-geometry/curvature**: Curvature as a measure of geodesic deviation — perturbation amplification in geometric terms
+
+## AI Designs It May Inspire
+
+- **Perturbation Regularization**: Add input or weight perturbation terms to the loss to train models insensitive to small changes
+- **Condition Number Monitor**: Track the condition number of weight matrices or the Hessian in real time to warn of ill-conditioning
+- **Robustness Verification Layer**: Quantify maximum output shift using epsilon-delta bounds, providing formal robustness guarantees
+- **Adaptive Preconditioning**: Dynamically adjust optimizer step size based on the perturbation amplification factor
+
+## Reasoning Protocol
+
+1. **Identify the perturbation source**: Does the perturbation occur in the input, parameters, or structure? What is its magnitude epsilon?
+2. **Compute sensitivity**: Derive the Jacobian or condition number — the ratio of output shift Delta to perturbation epsilon
+3. **Classify the perturbation**: Regular perturbation (expansion converges order by order) vs. singular perturbation (leading-order term changes)
+4. **Establish perturbation bounds**: Use known theorems (Weyl, Davis-Kahan, Lipschitz constants) to provide rigorous upper bounds
+5. **Assess robustness**: Is the perturbation amplification within acceptable range? If not, what regularization is needed?
+
+## Acceptance Criteria
+
+- The perturbation source and its magnitude epsilon are clearly defined
+- Sensitivity or condition number has been computed or estimated
+- The perturbation type has been classified (regular or singular)
+- The output shift has a quantified upper bound (not merely a heuristic estimate)
+- The robustness conclusion is actionable — pass, fail, or specification of required regularization

package/lenses/perturbation.md

@@ -0,0 +1,51 @@
+# 🧠 微扰透镜
+
+> 小扰动的传播、稳定性、鲁棒性——系统对微小变化的敏感度决定了它的可靠性
+
+## 它是什么视角
+
+这是一种"压力测试者"的视角——对任何数学对象或系统设计，追问"如果输入/参数/结构发生微小变化，输出会偏移多少？偏移是否可控？"核心信念：一个良态系统的输出对扰动的响应应当是有界的、可预测的；如果一个设计在无穷小扰动下就崩溃，那它在现实中注定失败。条件数是扰动放大倍数的精确度量。
+
+## 适合诊断什么问题
+
+- 模型对输入噪声/对抗扰动的鲁棒性分析
+- 优化问题中参数微变导致解的跳变——问题是病态的吗？
+- 特征值/特征向量在矩阵微扰下的稳定性（Davis-Kahan 定理）
+- 渐近展开中正则扰动 vs 奇异扰动的识别
+- 数值算法的误差传播与条件数分析
+
+## 不适合什么问题
+
+- 扰动本身就是研究对象（如混沌系统中蝴蝶效应是特征而非缺陷）
+- 需要大变形/全局结构分析——微扰只关心局部邻域
+- 离散组合问题——微扰理论依赖连续性假设
+
+## 会路由到哪些知识域
+
+- **matrix-analysis/matrix-perturbation**：Weyl 不等式、Davis-Kahan 定理、特征值/奇异值的扰动界
+- **optimization/constrained-optimization**：KKT 条件的稳定性、扰动后最优解的灵敏度
+- **probability/concentration-inequality**：随机扰动下的集中现象、亚高斯尾部界
+- **differential-geometry/curvature**：曲率作为测地线偏离的度量——几何意义上的扰动放大
+
+## 可能产生哪些 AI 设计
+
+- **扰动正则化**：在损失中加入输入/权重扰动项，训练对微小变化不敏感的模型
+- **条件数监控器**：实时跟踪权重矩阵/Hessian 的条件数，预警病态
+- **鲁棒性验证层**：用 ε-δ 界量化模型输出的最大偏移，提供形式化鲁棒性保证
+- **自适应预条件**：根据扰动放大因子动态调整优化器步长
+
+## 推理协议
+
+1. **识别扰动源**：扰动发生在输入、参数还是结构上？扰动的量级 ε 是多少？
+2. **计算灵敏度**：求导/Jacobian/条件数——输出偏移 Δ 与扰动 ε 的比值 Δ/ε
+3. **判断扰动类型**：正则扰动（展开式逐阶收敛）vs 奇异扰动（主导项阶次改变）
+4. **给出扰动界**：用已知定理（Weyl、Davis-Kahan、Lipschitz 常数）给出严格上界
+5. **评估鲁棒性**：扰动放大是否在可接受范围内？若不可接受，如何正则化？
+
+## 验收标准
+
+- 扰动源与扰动量级 ε 已明确定义
+- 灵敏度/条件数已计算或估计
+- 扰动类型已判定（正则/奇异）
+- 输出偏移有量化上界（非仅启发式估计）
+- 鲁棒性结论可执行——通过/不通过/需要何种正则化

package/lenses/probabilistic.en.md

@@ -0,0 +1,48 @@
+# 🎲 Probabilistic Lens
+
+> Uncertainty is quantifiable — probability is quantified belief, data is quantified evidence
+
+## What Perspective It Offers
+
+The probabilistic perspective is a way of "making rational decisions under uncertainty": the world is full of uncertainty, and rational decision-making does not mean pursuing certainty but rather making choices that optimize expected value amid uncertainty. Bayesian updating is the core operation — a prior, upon encountering new data, becomes a posterior; the posterior, upon encountering further data, becomes the new prior. A critical distinction must be maintained: statistical significance is not the same as practical significance (p < 0.05 does not mean a large effect), and correlation is not causation (P(Y|X) is not P(Y|do(X))).
+
+## What Problems It Is Suited to Diagnose
+
+- Quantifying uncertainty — probability distributions, confidence intervals, and effect-size estimation
+- Bayesian inference — step-by-step belief updating through the prior–likelihood–posterior cycle
+- Hypothesis testing and experimental design — significance tests, power analysis, and sample-size planning
+- Causal effect estimation — distinguishing statistical association from causal effect
+
+## What Problems It Is Not Suited For
+
+- Deterministic problems — the answer is known or can be obtained by deduction; probabilistic tools are unnecessary
+- Extremely small sample sizes (n < 5) with no possibility of increase — report the data itself rather than inferences from it
+- Data that is entirely missing or severely corrupted — no meaningful statistical operation can be performed
+
+## Which Knowledge Domains It Routes To
+
+- `probability/entropy`: Kolmogorov axioms, distribution families, and the law of large numbers / CLT — the mathematical foundations of uncertainty
+- `probability/kl-divergence`: MLE, hypothesis testing, Bayesian inference, and regression modeling — extracting patterns from data
+- causal inference (no KB card yet, use critic): DAGs, the back-door criterion, and do-calculus — the reasoning framework from association to causation
+
+## What AI Designs It May Inspire
+
+- **Belief-Updating Engine**: Implements an automatic chain of Bayesian prior-to-posterior updates
+- **Experiment Design Advisor**: Automatically computes the minimum sample size based on effect size and statistical power
+- **Causal Identification Module**: Constructs a DAG from variable relationships and identifies back-door paths and adjustment sets
+
+## Reasoning Protocol
+
+1. **Define the Random Phenomenon**: Specify the random variables, sample space, event space, and probability measure
+2. **Select a Probabilistic Model**: Choose a distribution based on data type, sample size, and the underlying physical mechanism
+3. **Statistical Inference**: Point estimation / MLE, confidence intervals, hypothesis testing, or Bayesian updating
+4. **Causal Assessment** (where applicable): Construct a DAG, identify confounders, and apply the back-door criterion
+5. **Quantify Uncertainty**: Provide confidence or credible intervals, report effect sizes, and distinguish statistical significance from practical significance
+
+## Acceptance Criteria
+
+- Random variables and distributions have been explicitly defined
+- The inference method (frequentist or Bayesian) has been selected with stated rationale
+- The meaning of p-values has been correctly understood (not "degree of confidence")
+- Correlation and causation have been distinguished, and confounders have been checked
+- Effect sizes and uncertainty have been reported, not merely point estimates

package/lenses/probabilistic.md

@@ -0,0 +1,48 @@
+# 🎲 概率透镜
+
+> 不确定性是可量化的——概率是量化的信念，数据是量化的证据
+
+## 它是什么视角
+
+概率是一种"在不确定性中做理性决策"的眼光：世界充满不确定性，理性决策不是追求确定性，而是在不确定性中做出期望值最优的选择。贝叶斯更新是核心操作——先验遇新数据变成后验，后验遇更新数据又成为新先验。关键区分：统计显著不等于实际显著（p<0.05 不等于效应大），相关不等于因果（P(Y|X) 不等于 P(Y|do(X))）。
+
+## 适合诊断什么问题
+
+- 量化不确定性——概率分布、置信区间、效应量评估
+- 贝叶斯推断——先验-似然-后验的逐步信念更新
+- 假设检验与实验设计——显著性检验、功效分析、样本量规划
+- 因果效应估计——区分统计关联与因果效应
+
+## 不适合什么问题
+
+- 确定性问题——已知答案或可演绎推理得到，不需要概率工具
+- 样本量极小（n<5）且无法增加——应报告数据本身而非推断
+- 数据完全缺失或严重损坏——无法做有意义的统计操作
+
+## 会路由到哪些知识域
+
+- `probability/entropy`：Kolmogorov 公理、分布族、大数定律/CLT——不确定性的数学基础
+- `probability/kl-divergence`：MLE、假设检验、贝叶斯推断、回归建模——从数据中提取规律
+- 因果推断（暂无知识卡，由 critic 审查）：DAG、后门准则、do-演算——从关联到因果的推理框架
+
+## 可能产生哪些 AI 设计
+
+- **信念更新引擎**：实现贝叶斯先验→后验的自动更新链条
+- **实验设计顾问**：根据效应量与功效自动计算最小样本量
+- **因果识别器**：从变量关系构建 DAG，识别后门路径与调整集
+
+## 推理协议
+
+1. **定义随机现象**：明确随机变量、样本空间、事件域、概率测度
+2. **选择概率模型**：根据数据类型、样本量、物理机制选择分布
+3. **统计推断**：点估计/MLE、置信区间、假设检验或贝叶斯更新
+4. **因果评估**（如适用）：构建 DAG，识别混淆变量，应用后门准则
+5. **量化不确定性**：给出置信/可信区间，报告效应量，区分统计显著与实际显著
+
+## 验收标准
+
+- 随机变量与分布已明确定义
+- 推断方法（频率/贝叶斯）已选择并说明理由
+- p-value 的含义已正确理解（非"把握程度"）
+- 相关与因果已区分，混淆变量已检查
+- 效应量与不确定性已报告，非仅点估计

package/lenses/projection.en.md

@@ -0,0 +1,50 @@
+# 🧮 Projection & Decomposition Lens
+
+> Complex wholes can be orthogonally decomposed into independent components — conflicts are exposed in subspaces, signal and noise are separated under projection.
+
+## What Perspective It Offers
+
+This is a "separator's" perspective — projecting mixed wholes onto orthogonal subspaces, splitting shared information from conflicting information, signal from noise, global structure from local detail. The core conviction: any vector can be decomposed into a parallel component plus an orthogonal component, and this decomposition is the first step toward understanding conflicts, eliminating redundancy, and compressing information.
+
+## What Problems It Is Suited to Diagnose
+
+- Multiple information sources are mixed; need to separate shared components from specific components (multi-domain learning, multimodal fusion)
+- Representations contain conflicting gradients or contradictory signals (multi-task learning, adversarial training)
+- Need to reduce dimensions or compress while preserving key structure (KV-Cache compression, feature selection)
+- Need to eliminate redundancy or orthogonalize multiple objectives (decorrelation, diversity constraints)
+
+## What Problems It Is Not Suited For
+
+- The problem itself is an indecomposable whole (strongly coupled systems, chaotic dynamics)
+- The subspace assumption is too strong — data does not lie on a low-dimensional subspace
+- Scenarios that require retaining all information with zero loss
+
+## Which Knowledge Domains It Routes To
+
+- **matrix-analysis/projection**: Orthogonal projection matrices, Courant-Fischer variational characterization
+- **matrix-analysis/spectral-decomposition**: EVD/SVD, principal component analysis
+- **matrix-analysis/low-rank-approximation**: Eckart-Young theorem, truncated SVD
+- **optimization/constrained-optimization**: Orthogonality-constrained optimization, Stiefel manifold
+
+## What AI Designs It May Inspire
+
+- **Shared-Private Decomposition**: Project multi-domain representations into a shared subspace plus domain-specific orthogonal complements
+- **Orthogonal Gradient Projection**: In multi-task learning, project new-task gradients onto the orthogonal complement of old-task gradients
+- **Low-Rank KV-Cache Compression**: Project K/V into a low-dimensional subspace, truncating weak components
+- **Head Diversity Constraint**: Force different attention heads to project onto approximately orthogonal subspaces
+
+## Reasoning Protocol
+
+1. **Identify mixed sources**: Which information, gradients, or representations are mixed together? Where do conflicts arise?
+2. **Define subspaces**: How are the dimensions of the shared space versus the specific space determined? What is the effective rank?
+3. **Construct projection operators**: $P = AA^H$ (orthonormal basis) or $P = A(A^HA)^{-1}A^H$ (general basis)
+4. **Perform decomposition**: $x = Px + (I-P)x$; evaluate the contribution of each component separately
+5. **Verify orthogonality**: Are $P^2 = P$ and $P = P^H$ satisfied? Is the condition number well-controlled?
+
+## Acceptance Criteria
+
+- The subspace decomposition is explicitly defined (dimensions, basis vectors, projection matrices)
+- The contributions of both components after projection have been separately quantified
+- Orthogonality conditions have been verified ($P^2 = P$, numerical condition number checked)
+- Compression/discard decisions are supported by quantified information-loss evidence
+- The output includes actionable conclusions, not merely the decomposition process

package/lenses/projection.md

@@ -0,0 +1,50 @@
+# 🧮 投影与分解透镜 / Projection & Decomposition Lens
+
+> 复杂整体可以正交分解为独立分量——冲突在子空间中暴露，信号与噪声在投影下分离。
+
+## 它是什么视角
+
+这是一种"分离者"的视角——将混杂的整体投影到正交子空间，把共享信息与冲突信息、信号与噪声、全局结构与局部细节拆分开来。核心信念：任何向量都可以分解为平行分量 + 正交分量，这种分解是理解冲突、消除冗余、压缩信息的第一步。
+
+## 适合诊断什么问题
+
+- 多个信息源混杂，需要分离共享成分与特异成分（多域学习、多模态融合）
+- 表示中存在冲突梯度或矛盾信号（多任务学习、对抗训练）
+- 需要降维或压缩但保留关键结构（KV-Cache 压缩、特征选择）
+- 需要消除冗余或正交化多个目标（去相关、diversity 约束）
+
+## 不适合什么问题
+
+- 问题本身就是不可分解的整体（强耦合系统、混沌动力学）
+- 子空间假设过强——数据不在低维子空间上
+- 需要保留全部信息不能有任何损失的场景
+
+## 会路由到哪些知识域
+
+- **matrix-analysis/projection**：正交投影矩阵、Courant-Fischer 变分刻画
+- **matrix-analysis/spectral-decomposition**：EVD/SVD、主成分分析
+- **matrix-analysis/low-rank-approximation**：Eckart-Young 定理、截断 SVD
+- **optimization/constrained-optimization**：正交约束优化、Stiefel 流形
+
+## 可能产生哪些 AI 设计
+
+- **Shared-Private 分解**：将多域表示投影为共享子空间 + 域特异正交补
+- **正交梯度投影**：在多任务学习中，将新任务梯度投影到旧任务梯度的正交补
+- **低秩 KV-Cache 压缩**：将 K/V 投影到低维子空间，截断弱分量
+- **Head diversity 约束**：强制不同 attention head 投影到近似正交的子空间
+
+## 推理协议
+
+1. **识别混杂源**：哪些信息/梯度/表示被混在一起？冲突在哪里发生？
+2. **定义子空间**：共享空间 vs 特异空间的维度如何确定？有效秩是多少？
+3. **构造投影算子**：$P = AA^H$（正交基）或 $P = A(A^HA)^{-1}A^H$（一般基）
+4. **执行分解**：$x = Px + (I-P)x$，分别评估两个分量的贡献
+5. **验证正交性**：$P^2 = P$、$P = P^H$ 是否满足？条件数是否可控？
+
+## 验收标准
+
+- 子空间分解已显式定义（维度、基向量、投影矩阵）
+- 投影后两个分量的贡献已分别量化
+- 正交性条件已验证（$P^2 = P$、数值条件数已检查）
+- 压缩/丢弃决策有信息损失量化依据
+- 输出包含可执行结论，不只是分解过程