{ "date": "2026-10-26", "scope": "Displayed mathematics, source explanations, deterministic implementations. Not empirical paper reproduction or full translation proofreading.", "entries": [ { "id": "adam", "note": [ "Initialize m₀=v₀=0; t≥1 and 0≤β₁,β₂<1. Operations are elementwise. Bias correction of v assumes a stationary raw second moment. The demo reports absolute update magnitudes.", "初始化 m₀=v₀=0；t≥1，0≤β₁,β₂<1。运算按元素进行。v 的偏差修正依赖平稳原始二阶矩；实验报告更新幅度的绝对值。" ], "check": [ "Implement scalar SGD and Adam on f(x,y)=x²+100y² using the same start and equal gradient-evaluation budgets. Log loss, gradient norm, moment estimates and actual step norm. Sweep learning rates for both methods with identical trial counts. Add a second test where the gradient changes sign to reveal momentum lag. Preserve failed runs rather than plotting only the best trajectory.", "在 f(x,y)=x²+100y² 上实现 SGD 与 Adam，保持起点和梯度计算预算相同。记录损失、梯度范数、矩估计和实际步长范数。用相同试验次数为两者搜索学习率，再加入梯度变号测试以观察动量滞后。保留失败运行，不只画最佳轨迹。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/adam/", "date": "2026-10-26" }, { "id": "adamw", "note": [ "The decay-only product assumes zero adaptive update, including zero historical moment state. Zero current gradients alone are insufficient. Shrinkage assumes 0≤αλ≤1.", "仅衰减乘积要求自适应更新为零，包括历史动量为零；仅当前梯度为零不够。称为收缩时需 0≤αλ≤1。" ], "check": [ "First verify equality for D=I with a unit test. Then compare actual Adam+L2 and AdamW on a small classifier, keeping data order and initialization paired. Search learning rate and decay jointly with equal budgets. Record parameter norms, train/validation loss and excluded parameter groups. State whether biases and normalization parameters receive decay.", "先用单元测试验证 D=I 时相等。再在小分类器上比较真正的 Adam+L2 和 AdamW，配对数据顺序与初始化。用相同预算联合搜索学习率和衰减，记录参数范数、训练/验证损失和排除的参数组。说明 bias 和归一化参数是否衰减。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/adamw/", "date": "2026-10-26" }, { "id": "transformer-attention", "note": [ "Self-attention: X is N×d_model; W_Q,W_K are d_model×d_k; W_V is d_model×d_v. Scores are N×N and output N×d_v. The variance argument assumes independent, zero-mean, unit-variance coordinates. Cross-attention can have different query/key lengths.", "自注意力：X 为 N×d_model；W_Q、W_K 为 d_model×d_k；W_V 为 d_model×d_v。分数 N×N，输出 N×d_v。方差推导假设坐标独立、零均值、单位方差；交叉注意力可有不同查询和键长度。" ], "check": [ "Build a four-token single-head reference implementation. Assert row sums and causal zeros. Perturb a future token and verify that earlier outputs do not change. Compare gradients with finite differences on a tiny input. Only after those checks, add multiple heads and compare batched output with a per-head implementation.", "实现四 token、单头基准版本，断言行和与因果零值。扰动未来 token，验证早先输出不变；在小输入上用有限差分检查梯度。通过这些检查后再加多头，比较批量计算与逐头实现。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/transformer-attention/", "date": "2026-10-26" }, { "id": "flashattention", "note": [ "Initialize running max to −∞ and normalizer/accumulator to zero. Skip empty or fully masked tiles; an entirely masked row has no probability distribution. Tiling reduces materialized memory, not quadratic dense arithmetic.", "运行最大值初始化为 −∞，归一化量及累积量为零。跳过空块和完全遮罩块；全遮罩行无法定义概率分布。分块减少物化内存，不消除稠密计算的平方复杂度。" ], "check": [ "Implement dense and tiled softmax attention on a small CPU array. Compare outputs with tolerance, including large logits and causal boundaries. For a real GPU benchmark, warm up kernels, synchronize timing, fix dtype and head dimensions, record peak allocated memory and separate forward from backward. Report distributions of runtime, not a single cold-start measurement.", "在小 CPU 数组上实现稠密与分块 softmax 注意力，按容差比较输出，覆盖大 logits 和因果边界。真实 GPU 基准应预热、同步计时、固定 dtype 与 head 维度，记录峰值分配内存，并分开正向和反向。报告运行时间分布，而非一次冷启动数值。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/flashattention/", "date": "2026-10-26" }, { "id": "lora", "note": [ "G denotes ∂L/∂W′. B is d_out×r and A is r×d_in; retain α/r in both gradients and the forward path. The singular-value residual bound concerns Frobenius low-rank approximation, not downstream language accuracy.", "G 表示 ∂L/∂W′。B 为 d_out×r，A 为 r×d_in；前向和梯度均保留 α/r。奇异值残差界针对 Frobenius 低秩逼近，不是下游语言准确率。" ], "check": [ "Start with the site’s small matrix-fit experiment. Sweep rank while holding target matrix and seed fixed, then compare with the singular-value lower bound. In a real language-model study, additionally fix target modules, adapter scaling, token budget and validation selection. Compare rank increases against wider module coverage at similar trainable-parameter budgets.", "先运行本站小矩阵拟合实验，固定目标矩阵与种子扫描 rank，并与奇异值下界比较。真实语言模型研究还需固定目标模块、适配器缩放、token 预算和验证选择。在相近参数预算下，比较增大 rank 与覆盖更多模块。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/lora/", "date": "2026-10-26" }, { "id": "qlora", "note": [ "Equal non-weight overhead is a comparison assumption. Real quantization metadata, activation storage and kernels differ. GB is decimal; GiB is binary. Raw 4-bit storage is not measured training peak memory.", "相同非权重开销是比较假设。实际量化元数据、激活与内核不同；GB 为十进制，GiB 为二进制。4 位原始存储不是实测训练峰值。" ], "check": [ "Profile a small supported model at matched sequence length, microbatch and optimizer. Record allocated and reserved peak memory separately, including a complete backward/optimizer step. Compare 16-bit LoRA and quantized LoRA using the same data and adapter targets. Sweep sequence length to identify when activations dominate; record quality and throughput along with memory.", "选小型受支持模型，固定序列长度、microbatch 和优化器做 profile。分别记录 allocated 与 reserved 峰值，覆盖完整反向和优化器步骤。用同样数据、适配器目标比较 16-bit LoRA 与量化 LoRA。扫描序列长度以判断何时激活主导，同时记录质量、吞吐和内存。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/qlora/", "date": "2026-10-26" }, { "id": "dpo", "note": [ "β>0; use natural logs and sum response-token log probabilities with a frozen reference. The sigmoid is pairwise preference probability, not response-generation probability. Minimize the loss.", "β>0，使用自然对数，回答 token 的 log 概率求和，参考模型固定。sigmoid 是成对偏好概率，不是回答生成概率；目标是最小化损失。" ], "check": [ "Create three synthetic preference pairs and manually compute the loss before training anything. Test that swapping chosen and rejected negates Δ. For a real POC, freeze the reference checkpoint, hold out prompts rather than individual responses, and compare against the supervised baseline. Track held-out preference accuracy, generation quality, length and estimated KL drift.", "先创建三组合成偏好对，手算损失，再训练。测试交换优选与落选是否使 Δ 变号。真实 POC 中固定参考 checkpoint，按 prompt 留出数据而非按单个 response；与监督基线比较，跟踪留出偏好准确率、生成质量、长度和估计 KL 漂移。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/dpo/", "date": "2026-10-26" }, { "id": "contrastive-predictive-coding", "note": [ "The mutual-information bound uses population expected InfoNCE loss: positives from the joint and independent negatives from the corresponding marginal. A single row loss is not a mutual-information estimate. Temperature must be positive.", "互信息下界使用总体期望 InfoNCE 损失：正样本来自联合分布，独立负样本来自对应边缘分布。单行损失不是互信息估计；温度必须为正。" ], "check": [ "Check analytical score gradients against finite differences. Then deliberately add a false negative and inspect its repulsive gradient. For an encoder POC, split data by source identity before constructing pairs, document the negative sampler and evaluate frozen representations with a held-out probe. Compare downstream metrics, not only contrastive loss.", "先用有限差分检查分数梯度。再故意加入假负样本，观察其排斥梯度。编码器 POC 应在构造样本对之前按来源身份划分数据，记录负采样器，并用留出 probe 评价冻结表示。比较下游指标，而不只对比损失。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/contrastive-predictive-coding/", "date": "2026-10-26" }, { "id": "simclr", "note": [ "Use cosine similarity and τ>0. Exclude the anchor itself, include its positive in the denominator, and average over 2N anchor directions. Freeze features and train separate linear probes for h/z with matched splits and tuning budgets.", "使用余弦相似度且 τ>0；分母排除自身、包含正样本，对 2N 个锚点方向平均。冻结特征，对 h/z 分别训练线性分类器，匹配划分和调参预算。" ], "check": [ "Design an augmentation validity table for your domain before running experiments. Compare no projection head versus a nonlinear head under a matched compute budget. Evaluate h and z separately with separate trained linear classifiers on frozen features, with matched splits and tuning budgets. Include a label-preserving and a deliberately destructive augmentation, recording retrieval failures and class-specific accuracy.", "先为你的领域制作增强有效性表，再做实验。在相同计算预算下比较无投影头和非线性头。用冻结特征上的独立训练线性分类器，并匹配数据划分和调参预算，分别评价 h 与 z。加入保留标签的增强和故意破坏信息的增强，记录检索失败与逐类准确率。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/simclr/", "date": "2026-10-26" }, { "id": "confidence-calibration", "note": [ "ECE here uses top-class confidence; empty bins contribute zero. Temperature is fitted on held-out calibration data. A lower ECE need not improve classification accuracy.", "此处 ECE 使用最高类别置信度；空分箱贡献零。温度在独立校准集拟合。ECE 降低不一定提高分类准确率。" ], "check": [ "Use disjoint training, calibration and test sets. Fit one positive T by calibration-set NLL, freeze it and evaluate once on test data. Report NLL, Brier score, reliability diagrams and bin counts alongside accuracy. Repeat the diagnostic on a shifted dataset without refitting to see whether the original correction transfers.", "使用独立训练集、校准集与测试集。按校准集 NLL 拟合正温度 T，固定后在测试集评价。除准确率外，报告 NLL、Brier score、可靠性图与分箱样本数。再在偏移数据集上不重新拟合地评价，检查校正能否迁移。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/confidence-calibration/", "date": "2026-10-26" }, { "id": "deep-rl-reproducibility", "note": [ "s/√n assumes IID independent experimental units with finite variance. A normal 1.96 interval is approximate and fragile at small n; compare a Student-t interval. Episodes from one training seed are not independent training runs.", "s/√n 假设独立同分布实验单元且方差有限。1.96 正态区间是近似，小样本时应比较 t 区间；同一训练种子的多个 episode 不是独立训练运行。" ], "check": [ "Predeclare the metric, run budget and stopping rule. Keep code, environment, dataset and evaluation horizon fixed across methods. Run multiple seeds and preserve raw per-run results, failures and exclusions. Compare an effect-size interval against a practical threshold, and audit whether tuning effort favored one baseline.", "预先声明指标、运行预算与停止规则。方法间固定代码、环境、数据和评估时域。运行多个种子，保留逐次原始结果、失败与排除项。将效应量区间与实际阈值比较，并审查调参投入是否偏向一方。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/deep-rl-reproducibility/", "date": "2026-10-26" }, { "id": "show-your-work", "note": [ "F(s)^n requires IID scores sharing the same CDF; independent nonidentical trials use the product of their individual CDFs. n is a positive integer. Best-of-budget and single-run means answer different questions.", "F(s)^n 要求分数独立同分布、共享 CDF；独立但不同分布需各 CDF 相乘。n 为正整数。预算内最优值与单次均值回答不同问题。" ], "check": [ "Log every hyperparameter trial with validation score, wall time, hardware and outcome. Compare methods over several matched budgets and specify whether the search is random, adaptive or manual. Select using validation only, then evaluate the selected checkpoint on a locked test set. Include unsuccessful runs in the cost accounting.", "记录每次超参数试验的验证分数、耗时、硬件和结果。在多个相同预算下比较方法，明确随机、自适应或人工搜索。只用验证集选择，再在锁定测试集评估选中 checkpoint。成本核算包含失败运行。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/show-your-work/", "date": "2026-10-26" }, { "id": "llm-as-a-judge", "note": [ "Comparison count must be positive. Preserve ties, map swapped labels back to original identities, and separately report parse failures. Synthetic ratings do not establish a real judge’s reliability.", "比较次数必须为正。保留平局，把交换标签映射回原身份，并单独报告解析失败。合成评分不证明真实裁判可靠。" ], "check": [ "Create a blinded rubric and a human-labeled audit subset. Randomize answer order, include swapped pairs, fix judge version and preserve raw outputs. Track ties, parse failures, verbosity differences and correctness disagreements. Bootstrap by item rather than by repeated judge call. Evaluate whether observed differences change a real deployment decision.", "创建盲评规则与人工标注审计子集。随机答案顺序，加入交换对，固定 judge 版本并保存原始输出。跟踪平局、解析失败、冗长度差异及正确性分歧。按题目而非重复调用做 bootstrap。检验差异是否改变真实部署决策。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/llm-as-a-judge/", "date": "2026-10-26" }, { "id": "react", "note": [ "Tool budgets are nonnegative integers. Allowlist checks precede execution; exhausted budgets stop the loop. This deterministic mock tests orchestration, not an LLM’s planning competence.", "工具预算为非负整数。执行前检查白名单，预算耗尽即停止。确定性 mock 测试编排逻辑，不代表 LLM 规划能力。" ], "check": [ "Build a deterministic tool mock with success, empty result, timeout and malformed-output cases. Record action arguments, observation summaries, termination reason and side effects. For a model experiment, compare prompting methods with the same tools, task set and action budget. Score external task completion, cost and unsafe-action attempts separately.", "构建确定性工具 mock，覆盖成功、空结果、超时和畸形输出。记录动作参数、观察摘要、终止原因与副作用。真实模型实验应在相同工具、任务集和行动预算下比较提示方法，分别评分外部完成情况、成本与不安全行动尝试。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/react/", "date": "2026-10-26" }, { "id": "ppo", "note": [ "Maximize the clipped surrogate, or minimize its negative. Freeze old log probabilities and advantages during each policy update. The sampled old-policy probability must be nonzero; clipping is not a hard KL constraint.", "最大化裁剪替代目标，或最小化其负值。策略更新期间固定旧 log 概率和 advantage；采样动作的旧概率须非零。裁剪不是硬 KL 约束。" ], "check": [ "Unit-test both advantage signs and both clipping boundaries. For a small environment POC, save rollout log probabilities and terminal flags, calculate advantages with a fixed convention and compare to a no-clipping baseline. Keep environment interactions and evaluation episodes matched. Log reward, entropy, approximate KL and clipping fraction for every update.", "单元测试两种 advantage 符号及两个裁剪边界。小环境 POC 保存 rollout log 概率与终止标记，按固定约定计算 advantage，并与无裁剪基线比较。匹配环境交互量与评测 episode。每次更新记录奖励、熵、近似 KL 与裁剪比例。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/ppo/", "date": "2026-10-26" }, { "id": "rlhf", "note": [ "KL is between response distributions, averaged over the same prompt distribution. Two example policies have expected reward/KL (1,0.4) and (0.8,0.1), so β=1 gives 0.6 and 0.7. A sampled log ratio can be negative; its expectation is KL.", "KL 定义在回答分布之间，并在同一 prompt 分布上平均。例中两种策略的期望奖励/KL 为 (1,0.4)、(0.8,0.1)，β=1 得 0.6、0.7。单次 log 比值可为负，其期望才是 KL。" ], "check": [ "Keep a blinded, held-out prompt set and compare base, SFT and preference-trained checkpoints at equal decoding budgets. Audit correctness separately from preference. For reward-model validation, measure pairwise agreement on new prompts and inspect disagreements. Include examples where verbosity, confidence or refusal style changes without factual improvement.", "保留盲化留出 prompt 集，在相同解码预算下比较基座、SFT 和偏好训练 checkpoint。将正确性与偏好分开审计。验证奖励模型时，测量新 prompt 上的成对一致性并检查分歧。加入冗长、自信或拒答风格改变但事实未改善的样本。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/rlhf/", "date": "2026-10-26" }, { "id": "deepseek-r1", "note": [ "Group normalization uses population variance (divide by G). Equal rewards remove the centered relative signal, not necessarily all gradients of an objective containing KL. This is not a reproduction of the full GRPO training algorithm.", "组内归一化使用总体方差（除以 G）。奖励相同消除中心化相对信号，不一定消除含 KL 的完整目标的全部梯度；本实验不复现完整 GRPO 训练。" ], "check": [ "Choose a small verifiable task with uncontaminated held-out problems. Compare training recipes at matched rollout and optimization budgets, then match inference tokens and sampling counts. Audit verifier false positives manually, preserve generated trajectories and measure pass@1 separately from best-of-k. Add a distribution-shift test and report failures, not only aggregate gains.", "选择小型可验证任务与无污染留出问题。匹配 rollout 与优化预算比较训练配方，再匹配推理 token 和采样数。人工审计 verifier 假阳性，保存轨迹，分开测量 pass@1 与 best-of-k。加入分布偏移测试，报告失败而非只有总体提升。" ], "files": [ "paper-core.mjs" ], "scope": [ "Displayed formula, bilingual explanation and declared slider range. This code reproduces the scalar teaching calculation, not the paper training system.", "复核所示公式、中英解释及声明滑块范围。代码复现标量教学计算，不是论文训练系统。" ], "path": "papers/deepseek-r1/", "date": "2026-10-26" }, { "id": "mdp-bellman", "note": [ "Condition the Bellman expectation on s,a. Continuing discounted tasks use 0≤γ<1; terminal value is zero. Rewards use r_(t+1), received after action a_t.", "Bellman 期望以 s,a 为条件。持续折扣任务取 0≤γ<1；终止价值为零。统一 r_(t+1) 表示执行 a_t 后收到的奖励。" ], "check": [ "Delayed option is 4.5; γ=0 makes it zero.", "延迟选项为 4.5；γ=0 时变为零。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "rl/mdp-bellman/", "date": "2026-10-26" }, { "id": "exploration-credit", "note": [ "Policy invariance requires bounded potential, the same γ, and zero terminal potential (or equivalent absorbing-tail handling). Telescoping gives G′=G−Φ(s₀)+γ^TΦ(s_T); an action-dependent final term can change optimal policy.", "策略不变要求有界势函数、相同 γ、终止势能为零（或等价吸收态尾项处理）。望远镜求和 G′=G−Φ(s₀)+γ^TΦ(s_T)；若尾项依赖动作，会改变最优策略。" ], "check": [ "Original return is 2.8, shaped return is 2.8−3. A nonzero terminal potential breaks this equality.", "原始回报为 2.8，塑形后为 2.8−3；非零终止势能会破坏此等式。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "rl/exploration-credit/", "date": "2026-10-26" }, { "id": "policy-gradient", "note": [ "Use δ_t=r_(t+1)+γ(1−d_t)V(s_(t+1))−V(s_t), with d marking true termination, not rollout truncation. GAE is a finite masked sum through T−1. Baselines are action-independent; stop gradients through advantages. State the discounted occupancy normalization when deriving ∇J.", "使用 δ_t=r_(t+1)+γ(1−d_t)V(s_(t+1))−V(s_t)，d 表示真正终止，不是采样截断。GAE 为到 T−1 的有限遮罩求和。基线不依赖动作，advantage 停止梯度；推导 ∇J 时明确折扣占用分布的归一化。" ], "check": [ "Compare analytic 0.15 to central finite differences of J at θ=0.", "在 θ=0 用中心差分比较解析梯度 0.15。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "rl/policy-gradient/", "date": "2026-10-26" }, { "id": "ppo-updates", "note": [ "Maximize the clipped surrogate, or minimize its negative. Freeze old log probabilities and advantages during each policy update. The sampled old-policy probability must be nonzero; clipping is not a hard KL constraint.", "最大化裁剪替代目标，或最小化其负值。策略更新期间固定旧 log 概率和 advantage；采样动作的旧概率须非零。裁剪不是硬 KL 约束。" ], "check": [ "Negative advantages require min after multiplying, not clipping the final objective.", "负 advantage 必须先相乘再取 min，不能裁剪最终目标值。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "rl/ppo-updates/", "date": "2026-10-26" }, { "id": "rlhf-dpo", "note": [ "β>0; use natural logs and sum response-token log probabilities with a frozen reference. The sigmoid is pairwise preference probability, not response-generation probability. Minimize the loss. RLHF maximizes reward minus KL, while SFT/DPO minimize losses.", "β>0，使用自然对数，回答 token 的 log 概率求和，参考模型固定。sigmoid 是成对偏好概率，不是回答生成概率；目标是最小化损失。 RLHF 最大化奖励减 KL，SFT/DPO 最小化损失。" ], "check": [ "At delta=0 loss is ln2; increasing preferred log odds decreases this loss.", "Δ=0 时损失为 ln2；提高偏好回答的 log odds 应降低损失。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "rl/rlhf-dpo/", "date": "2026-10-26" }, { "id": "grpo-verifiable", "note": [ "Group normalization uses population variance (divide by G). Equal rewards remove the centered relative signal, not necessarily all gradients of an objective containing KL. This is not a reproduction of the full GRPO training algorithm.", "组内归一化使用总体方差（除以 G）。奖励相同消除中心化相对信号，不一定消除含 KL 的完整目标的全部梯度；本实验不复现完整 GRPO 训练。" ], "check": [ "These exact values omit epsilon. Equal rewards require an explicit zero-variance branch or stabilizer.", "此处精确值忽略 ε；全部奖励相同时必须处理零方差。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "rl/grpo-verifiable/", "date": "2026-10-26" }, { "id": "agent-rl", "note": [ "Tool budgets are nonnegative integers. Allowlist checks precede execution; exhausted budgets stop the loop. This deterministic mock tests orchestration, not an LLM’s planning competence.", "工具预算为非负整数。执行前检查白名单，预算耗尽即停止。确定性 mock 测试编排逻辑，不代表 LLM 规划能力。" ], "check": [ "The third action is never executed. This tests a budget gate, not a learned agent.", "第三个动作不得执行。这里测试预算关卡，不是学习型 Agent。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "rl/agent-rl/", "date": "2026-10-26" }, { "id": "rl-evaluation", "note": [ "s/√n assumes IID independent experimental units with finite variance. A normal 1.96 interval is approximate and fragile at small n; compare a Student-t interval. Episodes from one training seed are not independent training runs.", "s/√n 假设独立同分布实验单元且方差有限。1.96 正态区间是近似，小样本时应比较 t 区间；同一训练种子的多个 episode 不是独立训练运行。" ], "check": [ "Sample variance is 2.5 and SE is sqrt(0.5); episodes cannot replace independent seeds.", "样本方差为 2.5，标准误为 sqrt(0.5)；episode 不能替代独立种子。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "rl/rl-evaluation/", "date": "2026-10-26" }, { "id": "frames-and-kinematics", "note": [ "Rᵀ is the inverse only for orthonormal R; proper rigid rotations also have det(R)=1. Use right-handed frames and matching length units; homogeneous transforms act on column vectors.", "Rᵀ 为逆要求 R 正交；刚体旋转还需 det(R)=1。采用右手坐标、统一长度单位，齐次变换作用于列向量。" ], "check": [ "Round trip recovers p. A nonorthogonal R cannot use transpose as inverse.", "往返变换恢复 p；非正交 R 不能用转置代替逆。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "physical-ai/frames-and-kinematics/", "date": "2026-10-26" }, { "id": "perception-and-state", "note": [ "The scalar Kalman example assumes direct measurement H=1, independent zero-mean noises and positive innovation variance. Covariances use squared position units.", "标量 Kalman 算例假设直接测量 H=1、独立零均值噪声和正创新方差；协方差单位是位置单位的平方。" ], "check": [ "Posterior lies between prior and measurement; posterior variance is nonnegative and smaller.", "后验均值位于先验和测量之间，后验方差非负且变小。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "physical-ai/perception-and-state/", "date": "2026-10-26" }, { "id": "feedback-and-planning", "note": [ "Kp=10 N/m and Kd=2 N·s/m. Quadratic MPC uses Q positive semidefinite and R positive definite; convex optimization additionally needs affine dynamics and convex constraints. The example uses explicit Euler.", "Kp=10 牛顿/米，Kd=2 牛顿·秒/米。二次 MPC 常取 Q 半正定、R 正定；凸优化还需仿射动力学和凸约束。算例采用显式 Euler。" ], "check": [ "Explicit Euler updates position from old velocity. Changing to new velocity is a different integrator.", "显式 Euler 使用旧速度更新位置；使用新速度是另一种积分器。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "physical-ai/feedback-and-planning/", "date": "2026-10-26" }, { "id": "imitation-and-act", "note": [ "Action chunks are supervised sequences; temporal ensembling blends overlapping predictions for the same execution time. A low offline regression loss does not eliminate closed-loop distribution shift.", "动作块是监督序列；时间集成混合同一执行时刻的重叠预测。低离线回归损失不能消除闭环分布偏移。" ], "check": [ "Predictions must refer to the same execution time; normalized nonnegative weights form a convex combination.", "预测必须对应同一执行时刻；归一化非负权重构成凸组合。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "physical-ai/imitation-and-act/", "date": "2026-10-26" }, { "id": "diffusion-policy", "note": [ "Noise is added in normalized action coordinates. The forward noise example is not a reverse sampler or a motor command. Enforce physical constraints after decoding, and evaluate closed-loop interventions.", "加噪在规范化动作坐标中进行。前向加噪算例不是反向采样器，也不是电机指令。解码后仍需物理约束，并评价闭环干预。" ], "check": [ "alphaBar=1 recovers clean normalized action; alphaBar=0 gives noise. This is forward diffusion only.", "ᾱ=1 恢复干净动作，ᾱ=0 得到噪声；这里只计算前向扩散。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "physical-ai/diffusion-policy/", "date": "2026-10-26" }, { "id": "vla-openvla", "note": [ "Action bins quantize normalized coordinates. Half-bin error assumes nearest-bin rounding away from boundary/clipping effects. Token prediction alone does not guarantee calibration, control stability or safety.", "动作分箱量化规范化坐标。半格误差界依赖最近格点舍入，并需考虑边界/裁剪。预测 token 不保证标定、控制稳定或安全。" ], "check": [ "For values within bounds nearest-bin absolute error is at most half a step.", "在取值范围内，最近格点舍入的绝对误差不超过半个步长。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "physical-ai/vla-openvla/", "date": "2026-10-26" }, { "id": "flow-matching-pi0", "note": [ "t is dimensionless interpolation time. vθ has action-coordinate units per interpolation time, not physical robot velocity. The squared-error optimum is E[a−ε|x_t,t,o]; generative x_next is not the next physical robot state.", "t 是无量纲插值时间。vθ 单位是动作坐标/插值时间，不是机器人真实速度。平方误差最优解为 E[a−ε|x_t,t,o]；生成状态 x_next 不是下一物理状态。" ], "check": [ "4 is velocity in normalized action/interpolation-time coordinates, not 4 m/s. This is one paired target.", "4 是规范化动作/插值时间的速度，不是 4 米/秒；这里只是一个配对目标。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "physical-ai/flow-matching-pi0/", "date": "2026-10-26" }, { "id": "rl-sim2real-evaluation", "note": [ "Zero failures does not imply zero risk. The one-sided 95% upper bound 1−0.05^(1/n) assumes independent Bernoulli trials; deployment drift and correlated resets violate that model.", "零失败不等于零风险。单侧 95% 上界 1−0.05^(1/n) 假设独立 Bernoulli 试验；部署漂移及相关重置会违反该模型。" ], "check": [ "Zero failures in 20 IID trials still permits about 13.9% failure probability at the one-sided 95% bound.", "20 次独立同分布试验零失败，单侧 95% 失败率上界仍约为 13.9%。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "physical-ai/rl-sim2real-evaluation/", "date": "2026-10-26" }, { "id": "kv-serving", "note": [ "Count both K and V, KV heads rather than query heads, and sequence×batch exactly once. The byte formula excludes activations, weights, allocator fragmentation and attention workspace.", "同时计 K/V，使用 KV 头数而不是 query 头数；序列×批次仅乘一次。字节公式不含激活、权重、分配碎片和注意力工作区。" ], "check": [ "Double batch doubles MiB; change query heads alone and KV storage stays unchanged.", "批次翻倍则 MiB 翻倍；仅修改 query 头数不改变 KV 存储。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "systems/kv-serving/", "date": "2026-10-26" }, { "id": "speculative-correctness", "note": [ "Conditional acceptance is min(1,p(x)/q(x)) for q(x)>0. Joint proposal-and-acceptance mass is min(p(x),q(x)). Rejection samples normalized positive residual; identical distributions need no residual draw.", "q(x)>0 时条件接受率为 min(1,p(x)/q(x))；提出且接受 x 的联合质量为 min(p(x),q(x))。拒绝后采样归一化正残差；两分布相同无需残差采样。" ], "check": [ "Accepted mass is .7, rejection mass .3; adding the residual exactly restores target p.", "接受质量为 0.7，拒绝质量为 0.3；补回残差后必须精确恢复 p。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "systems/speculative-correctness/", "date": "2026-10-26" }, { "id": "moe-budgets", "note": [ "Activated parameter count is not total resident weight memory. Routing capacity, token imbalance, communication and expert precision all affect runtime. Distinguish logical dispatch bytes from measured network traffic.", "激活参数量不是驻留权重总内存。路由容量、token 不均衡、通信和专家精度均影响运行；逻辑分派字节不同于实测网络流量。" ], "check": [ "256 expert assignments dispatch 2 MiB once; return traffic is additional. This excludes protocol overhead.", "256 次专家分派单向传输 2 MiB；回传另计，且未计协议开销。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "systems/moe-budgets/", "date": "2026-10-26" }, { "id": "test-time-compute", "note": [ "C_verification(N) is total verification cost for N candidates. Oracle coverage assumes independent attempts with equal success probability; selected accuracy further depends on the verifier.", "C_verification(N) 是 N 个候选的总验证成本。oracle 覆盖率假设独立等成功率尝试；最终准确率还取决于验证器。" ], "check": [ "Perfectly correlated attempts retain .2 coverage; the IID formula cannot be reused.", "完全相关的尝试仍只有 0.2 覆盖率；不能套用独立同分布公式。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "frontier/test-time-compute/", "date": "2026-10-26" }, { "id": "selective-state-spaces", "note": [ "A scalar recurrence illustrates state compression, not the full selective SSM architecture. State size and numerical discretization matter; linear recurrence is not equivalent to unrestricted attention retrieval.", "标量递推用于展示状态压缩，不是完整选择性 SSM。状态大小及数值离散化重要；线性递推不等价于无限制注意力检索。" ], "check": [ "Impulse memory decays geometrically; a=1 preserves it and |a|>1 amplifies it.", "脉冲记忆几何衰减；a=1 保留，|a|>1 放大。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "frontier/selective-state-spaces/", "date": "2026-10-26" }, { "id": "diffusion-language", "note": [ "The displayed cross-entropy uses natural logarithms and is measured in nats. A masked-position average must divide by the number of evaluated masks; the toy loss does not reproduce an entire discrete diffusion objective.", "所示交叉熵使用自然对数，单位 nats。遮罩位置均值须除以实际评价的遮罩数；教学损失不复现完整离散扩散目标。" ], "check": [ "The result is nats. Divide by ln2 for bits; do not divide by unmasked token count.", "结果单位为 nats；除以 ln2 得 bits，不能按未遮罩 token 数平均。" ], "files": [], "scope": [ "Formula conditions, units and worked arithmetic reviewed. The implementation isolates one numerical mechanism; it does not train a policy or control real hardware.", "已复核公式条件、单位及算例。实现隔离一个数值机制，不训练完整策略，也不控制真实硬件。" ], "path": "frontier/diffusion-language/", "date": "2026-10-26" }, { "id": "attention-1", "note": [ "Self-attention: X is N×d_model; W_Q,W_K are d_model×d_k; W_V is d_model×d_v. Scores are N×N and output N×d_v. The variance argument assumes independent, zero-mean, unit-variance coordinates. Cross-attention can have different query/key lengths.", "自注意力：X 为 N×d_model；W_Q、W_K 为 d_model×d_k；W_V 为 d_model×d_v。分数 N×N，输出 N×d_v。方差推导假设坐标独立、零均值、单位方差；交叉注意力可有不同查询和键长度。" ], "check": [ "Future positions must receive zero probability.", "Future positions must receive zero probability." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/attention-1/", "date": "2026-10-26" }, { "id": "attention-2", "note": [ "Self-attention: X is N×d_model; W_Q,W_K are d_model×d_k; W_V is d_model×d_v. Scores are N×N and output N×d_v. The variance argument assumes independent, zero-mean, unit-variance coordinates. Cross-attention can have different query/key lengths.", "自注意力：X 为 N×d_model；W_Q、W_K 为 d_model×d_k；W_V 为 d_model×d_v。分数 N×N，输出 N×d_v。方差推导假设坐标独立、零均值、单位方差；交叉注意力可有不同查询和键长度。" ], "check": [ "Removing the mask exposes future keys.", "Removing the mask exposes future keys." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/attention-2/", "date": "2026-10-26" }, { "id": "attention-3", "note": [ "Self-attention: X is N×d_model; W_Q,W_K are d_model×d_k; W_V is d_model×d_v. Scores are N×N and output N×d_v. The variance argument assumes independent, zero-mean, unit-variance coordinates. Cross-attention can have different query/key lengths.", "自注意力：X 为 N×d_model；W_Q、W_K 为 d_model×d_k；W_V 为 d_model×d_v。分数 N×N，输出 N×d_v。方差推导假设坐标独立、零均值、单位方差；交叉注意力可有不同查询和键长度。" ], "check": [ "Larger logits concentrate probability on the largest visible score.", "Larger logits concentrate probability on the largest visible score." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/attention-3/", "date": "2026-10-26" }, { "id": "attention-4", "note": [ "Self-attention: X is N×d_model; W_Q,W_K are d_model×d_k; W_V is d_model×d_v. Scores are N×N and output N×d_v. The variance argument assumes independent, zero-mean, unit-variance coordinates. Cross-attention can have different query/key lengths.", "自注意力：X 为 N×d_model；W_Q、W_K 为 d_model×d_k；W_V 为 d_model×d_v。分数 N×N，输出 N×d_v。方差推导假设坐标独立、零均值、单位方差；交叉注意力可有不同查询和键长度。" ], "check": [ "Equal visible logits yield uniform weights.", "Equal visible logits yield uniform weights." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/attention-4/", "date": "2026-10-26" }, { "id": "attention-5", "note": [ "Self-attention: X is N×d_model; W_Q,W_K are d_model×d_k; W_V is d_model×d_v. Scores are N×N and output N×d_v. The variance argument assumes independent, zero-mean, unit-variance coordinates. Cross-attention can have different query/key lengths.", "自注意力：X 为 N×d_model；W_Q、W_K 为 d_model×d_k；W_V 为 d_model×d_v。分数 N×N，输出 N×d_v。方差推导假设坐标独立、零均值、单位方差；交叉注意力可有不同查询和键长度。" ], "check": [ "A smaller scale moves weights toward uniform without changing visibility.", "A smaller scale moves weights toward uniform without changing visibility." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/attention-5/", "date": "2026-10-26" }, { "id": "embedding-1", "note": [ "Mathematical cosine is undefined for a zero vector. This software returns zero as a sentinel, which must not be interpreted as orthogonality. Dot product and cosine have different scale sensitivity.", "零向量的数学余弦未定义。程序返回零作为哨兵值，不能解读为正交。点积与余弦的尺度敏感性不同。" ], "check": [ "A large document norm should not dominate cosine ranking.", "A large document norm should not dominate cosine ranking." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/embedding-1/", "date": "2026-10-26" }, { "id": "embedding-2", "note": [ "Mathematical cosine is undefined for a zero vector. This software returns zero as a sentinel, which must not be interpreted as orthogonality. Dot product and cosine have different scale sensitivity.", "零向量的数学余弦未定义。程序返回零作为哨兵值，不能解读为正交。点积与余弦的尺度敏感性不同。" ], "check": [ "The long vector may outrank a better-aligned short vector.", "The long vector may outrank a better-aligned short vector." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/embedding-2/", "date": "2026-10-26" }, { "id": "embedding-3", "note": [ "Mathematical cosine is undefined for a zero vector. This software returns zero as a sentinel, which must not be interpreted as orthogonality. Dot product and cosine have different scale sensitivity.", "零向量的数学余弦未定义。程序返回零作为哨兵值，不能解读为正交。点积与余弦的尺度敏感性不同。" ], "check": [ "Rotating the query changes which document aligns best.", "Rotating the query changes which document aligns best." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/embedding-3/", "date": "2026-10-26" }, { "id": "embedding-4", "note": [ "Mathematical cosine is undefined for a zero vector. This software returns zero as a sentinel, which must not be interpreted as orthogonality. Dot product and cosine have different scale sensitivity.", "零向量的数学余弦未定义。程序返回零作为哨兵值，不能解读为正交。点积与余弦的尺度敏感性不同。" ], "check": [ "Opposite vectors receive negative cosine scores.", "Opposite vectors receive negative cosine scores." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/embedding-4/", "date": "2026-10-26" }, { "id": "embedding-5", "note": [ "Mathematical cosine is undefined for a zero vector. This software returns zero as a sentinel, which must not be interpreted as orthogonality. Dot product and cosine have different scale sensitivity.", "零向量的数学余弦未定义。程序返回零作为哨兵值，不能解读为正交。点积与余弦的尺度敏感性不同。" ], "check": [ "The lab returns zero scores for an undefined cosine denominator.", "The lab returns zero scores for an undefined cosine denominator." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/embedding-5/", "date": "2026-10-26" }, { "id": "kv-1", "note": [ "Count both K and V, KV heads rather than query heads, and sequence×batch exactly once. The byte formula excludes activations, weights, allocator fragmentation and attention workspace.", "同时计 K/V，使用 KV 头数而不是 query 头数；序列×批次仅乘一次。字节公式不含激活、权重、分配碎片和注意力工作区。" ], "check": [ "Eight KV heads use one quarter the cache of 32 heads.", "Eight KV heads use one quarter the cache of 32 heads." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/kv-1/", "date": "2026-10-26" }, { "id": "kv-2", "note": [ "Count both K and V, KV heads rather than query heads, and sequence×batch exactly once. The byte formula excludes activations, weights, allocator fragmentation and attention workspace.", "同时计 K/V，使用 KV 头数而不是 query 头数；序列×批次仅乘一次。字节公式不含激活、权重、分配碎片和注意力工作区。" ], "check": [ "More KV heads scale cache bytes linearly.", "More KV heads scale cache bytes linearly." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/kv-2/", "date": "2026-10-26" }, { "id": "kv-3", "note": [ "Count both K and V, KV heads rather than query heads, and sequence×batch exactly once. The byte formula excludes activations, weights, allocator fragmentation and attention workspace.", "同时计 K/V，使用 KV 头数而不是 query 头数；序列×批次仅乘一次。字节公式不含激活、权重、分配碎片和注意力工作区。" ], "check": [ "One KV head sharply reduces logical cache size.", "One KV head sharply reduces logical cache size." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/kv-3/", "date": "2026-10-26" }, { "id": "kv-4", "note": [ "Count both K and V, KV heads rather than query heads, and sequence×batch exactly once. The byte formula excludes activations, weights, allocator fragmentation and attention workspace.", "同时计 K/V，使用 KV 头数而不是 query 头数；序列×批次仅乘一次。字节公式不含激活、权重、分配碎片和注意力工作区。" ], "check": [ "Sequence and batch size multiply memory demand.", "Sequence and batch size multiply memory demand." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/kv-4/", "date": "2026-10-26" }, { "id": "kv-5", "note": [ "Count both K and V, KV heads rather than query heads, and sequence×batch exactly once. The byte formula excludes activations, weights, allocator fragmentation and attention workspace.", "同时计 K/V，使用 KV 头数而不是 query 头数；序列×批次仅乘一次。字节公式不含激活、权重、分配碎片和注意力工作区。" ], "check": [ "One extra token allocates an additional page per sequence.", "One extra token allocates an additional page per sequence." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/kv-5/", "date": "2026-10-26" }, { "id": "neural-1", "note": [ "XOR labels are 1 exactly when the signs differ. Stable logits BCE is max(z,0)−zy+log1p(exp(−|z|)); its logit gradient is sigmoid(z)−y. Training and held-out sets are separately seeded samples, not real-world validation.", "XOR 标签恰在符号不同时为 1。稳定 logits BCE 为 max(z,0)−zy+log1p(exp(−|z|))，logit 梯度为 sigmoid(z)−y。训练及留出集是独立生成的合成样本，不是现实验证。" ], "check": [ "A nonlinear hidden layer can fit this non-linearly separable task.", "A nonlinear hidden layer can fit this non-linearly separable task." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/neural-1/", "date": "2026-10-26" }, { "id": "neural-2", "note": [ "XOR labels are 1 exactly when the signs differ. Stable logits BCE is max(z,0)−zy+log1p(exp(−|z|)); its logit gradient is sigmoid(z)−y. Training and held-out sets are separately seeded samples, not real-world validation.", "XOR 标签恰在符号不同时为 1。稳定 logits BCE 为 max(z,0)−zy+log1p(exp(−|z|))，logit 梯度为 sigmoid(z)−y。训练及留出集是独立生成的合成样本，不是现实验证。" ], "check": [ "One hidden unit limits the available decision boundaries.", "One hidden unit limits the available decision boundaries." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/neural-2/", "date": "2026-10-26" }, { "id": "neural-3", "note": [ "XOR labels are 1 exactly when the signs differ. Stable logits BCE is max(z,0)−zy+log1p(exp(−|z|)); its logit gradient is sigmoid(z)−y. Training and held-out sets are separately seeded samples, not real-world validation.", "XOR 标签恰在符号不同时为 1。稳定 logits BCE 为 max(z,0)−zy+log1p(exp(−|z|))，logit 梯度为 sigmoid(z)−y。训练及留出集是独立生成的合成样本，不是现实验证。" ], "check": [ "A lower learning rate may need more epochs.", "A lower learning rate may need more epochs." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/neural-3/", "date": "2026-10-26" }, { "id": "neural-4", "note": [ "XOR labels are 1 exactly when the signs differ. Stable logits BCE is max(z,0)−zy+log1p(exp(−|z|)); its logit gradient is sigmoid(z)−y. Training and held-out sets are separately seeded samples, not real-world validation.", "XOR 标签恰在符号不同时为 1。稳定 logits BCE 为 max(z,0)−zy+log1p(exp(−|z|))，logit 梯度为 sigmoid(z)−y。训练及留出集是独立生成的合成样本，不是现实验证。" ], "check": [ "The hidden layer can approximate a curved boundary.", "The hidden layer can approximate a curved boundary." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/neural-4/", "date": "2026-10-26" }, { "id": "neural-5", "note": [ "XOR labels are 1 exactly when the signs differ. Stable logits BCE is max(z,0)−zy+log1p(exp(−|z|)); its logit gradient is sigmoid(z)−y. Training and held-out sets are separately seeded samples, not real-world validation.", "XOR 标签恰在符号不同时为 1。稳定 logits BCE 为 max(z,0)−zy+log1p(exp(−|z|))，logit 梯度为 sigmoid(z)−y。训练及留出集是独立生成的合成样本，不是现实验证。" ], "check": [ "Initialization and sampled data affect the learning trajectory.", "Initialization and sampled data affect the learning trajectory." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/neural-5/", "date": "2026-10-26" }, { "id": "agent-1", "note": [ "Tool budgets are nonnegative integers. Allowlist checks precede execution; exhausted budgets stop the loop. This deterministic mock tests orchestration, not an LLM’s planning competence.", "工具预算为非负整数。执行前检查白名单，预算耗尽即停止。确定性 mock 测试编排逻辑，不代表 LLM 规划能力。" ], "check": [ "A write pauses without the mock approval toggle.", "A write pauses without the mock approval toggle." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/agent-1/", "date": "2026-10-26" }, { "id": "agent-2", "note": [ "Tool budgets are nonnegative integers. Allowlist checks precede execution; exhausted budgets stop the loop. This deterministic mock tests orchestration, not an LLM’s planning competence.", "工具预算为非负整数。执行前检查白名单，预算耗尽即停止。确定性 mock 测试编排逻辑，不代表 LLM 规划能力。" ], "check": [ "Approval permits only the local mock write.", "Approval permits only the local mock write." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/agent-2/", "date": "2026-10-26" }, { "id": "agent-3", "note": [ "Tool budgets are nonnegative integers. Allowlist checks precede execution; exhausted budgets stop the loop. This deterministic mock tests orchestration, not an LLM’s planning competence.", "工具预算为非负整数。执行前检查白名单，预算耗尽即停止。确定性 mock 测试编排逻辑，不代表 LLM 规划能力。" ], "check": [ "A disallowed calculator call stops the workflow.", "A disallowed calculator call stops the workflow." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/agent-3/", "date": "2026-10-26" }, { "id": "agent-4", "note": [ "Tool budgets are nonnegative integers. Allowlist checks precede execution; exhausted budgets stop the loop. This deterministic mock tests orchestration, not an LLM’s planning competence.", "工具预算为非负整数。执行前检查白名单，预算耗尽即停止。确定性 mock 测试编排逻辑，不代表 LLM 规划能力。" ], "check": [ "The second paid action is blocked before exceeding budget.", "The second paid action is blocked before exceeding budget." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/agent-4/", "date": "2026-10-26" }, { "id": "agent-5", "note": [ "Tool budgets are nonnegative integers. Allowlist checks precede execution; exhausted budgets stop the loop. This deterministic mock tests orchestration, not an LLM’s planning competence.", "工具预算为非负整数。执行前检查白名单，预算耗尽即停止。确定性 mock 测试编排逻辑，不代表 LLM 规划能力。" ], "check": [ "Only the latest result survives FIFO truncation.", "Only the latest result survives FIFO truncation." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/agent-5/", "date": "2026-10-26" }, { "id": "retrieval-1", "note": [ "Weighted reciprocal-rank fusion uses ranks starting at 1. Supplied relevance is an oracle reranker, not a trained model. Zero-vector cosine uses a software sentinel; labels must not enter a real retrieval index.", "加权倒数排名融合从排名 1 开始。给定相关性是 oracle 重排器，不是训练模型。零向量余弦采用软件哨兵值；实际检索索引不得泄漏标签。" ], "check": [ "Combining independent ranks can promote shared candidates.", "Combining independent ranks can promote shared candidates." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/retrieval-1/", "date": "2026-10-26" }, { "id": "retrieval-2", "note": [ "Weighted reciprocal-rank fusion uses ranks starting at 1. Supplied relevance is an oracle reranker, not a trained model. Zero-vector cosine uses a software sentinel; labels must not enter a real retrieval index.", "加权倒数排名融合从排名 1 开始。给定相关性是 oracle 重排器，不是训练模型。零向量余弦采用软件哨兵值；实际检索索引不得泄漏标签。" ], "check": [ "Exact query words alone set the ranking.", "Exact query words alone set the ranking." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/retrieval-2/", "date": "2026-10-26" }, { "id": "retrieval-3", "note": [ "Weighted reciprocal-rank fusion uses ranks starting at 1. Supplied relevance is an oracle reranker, not a trained model. Zero-vector cosine uses a software sentinel; labels must not enter a real retrieval index.", "加权倒数排名融合从排名 1 开始。给定相关性是 oracle 重排器，不是训练模型。零向量余弦采用软件哨兵值；实际检索索引不得泄漏标签。" ], "check": [ "Only vector direction determines the ranking.", "Only vector direction determines the ranking." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/retrieval-3/", "date": "2026-10-26" }, { "id": "retrieval-4", "note": [ "Weighted reciprocal-rank fusion uses ranks starting at 1. Supplied relevance is an oracle reranker, not a trained model. Zero-vector cosine uses a software sentinel; labels must not enter a real retrieval index.", "加权倒数排名融合从排名 1 开始。给定相关性是 oracle 重排器，不是训练模型。零向量余弦采用软件哨兵值；实际检索索引不得泄漏标签。" ], "check": [ "Perfect labels can improve ordering only inside the candidate pool.", "Perfect labels can improve ordering only inside the candidate pool." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/retrieval-4/", "date": "2026-10-26" }, { "id": "retrieval-5", "note": [ "Weighted reciprocal-rank fusion uses ranks starting at 1. Supplied relevance is an oracle reranker, not a trained model. Zero-vector cosine uses a software sentinel; labels must not enter a real retrieval index.", "加权倒数排名融合从排名 1 开始。给定相关性是 oracle 重排器，不是训练模型。零向量余弦采用软件哨兵值；实际检索索引不得泄漏标签。" ], "check": [ "Even an oracle cannot recover documents excluded before reranking.", "Even an oracle cannot recover documents excluded before reranking." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/retrieval-5/", "date": "2026-10-26" }, { "id": "judge-1", "note": [ "Comparison count must be positive. Preserve ties, map swapped labels back to original identities, and separately report parse failures. Synthetic ratings do not establish a real judge’s reliability.", "比较次数必须为正。保留平局，把交换标签映射回原身份，并单独报告解析失败。合成评分不证明真实裁判可靠。" ], "check": [ "Agreement and order consistency measure different properties.", "Agreement and order consistency measure different properties." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/judge-1/", "date": "2026-10-26" }, { "id": "judge-2", "note": [ "Comparison count must be positive. Preserve ties, map swapped labels back to original identities, and separately report parse failures. Synthetic ratings do not establish a real judge’s reliability.", "比较次数必须为正。保留平局，把交换标签映射回原身份，并单独报告解析失败。合成评分不证明真实裁判可靠。" ], "check": [ "Perfect labels yield agreement and consistency of one.", "Perfect labels yield agreement and consistency of one." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/judge-2/", "date": "2026-10-26" }, { "id": "judge-3", "note": [ "Comparison count must be positive. Preserve ties, map swapped labels back to original identities, and separately report parse failures. Synthetic ratings do not establish a real judge’s reliability.", "比较次数必须为正。保留平局，把交换标签映射回原身份，并单独报告解析失败。合成评分不证明真实裁判可靠。" ], "check": [ "Selecting the first position fails swap consistency.", "Selecting the first position fails swap consistency." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/judge-3/", "date": "2026-10-26" }, { "id": "judge-4", "note": [ "Comparison count must be positive. Preserve ties, map swapped labels back to original identities, and separately report parse failures. Synthetic ratings do not establish a real judge’s reliability.", "比较次数必须为正。保留平局，把交换标签映射回原身份，并单独报告解析失败。合成评分不证明真实裁判可靠。" ], "check": [ "A large tie class can make aggregate agreement misleading.", "A large tie class can make aggregate agreement misleading." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/judge-4/", "date": "2026-10-26" }, { "id": "judge-5", "note": [ "Comparison count must be positive. Preserve ties, map swapped labels back to original identities, and separately report parse failures. Synthetic ratings do not establish a real judge’s reliability.", "比较次数必须为正。保留平局，把交换标签映射回原身份，并单独报告解析失败。合成评分不证明真实裁判可靠。" ], "check": [ "A judge can be order-consistent and disagree with every human label.", "A judge can be order-consistent and disagree with every human label." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/judge-5/", "date": "2026-10-26" }, { "id": "routing-1", "note": [ "Failed cheap calls are billed and fallback is assumed successful. Expected latency is serial and excludes queues; aggregate expectation is not p95 latency.", "失败的便宜调用仍计费，假设 fallback 成功。期望延迟按串行计算且忽略排队；总体期望不是 p95 延迟。" ], "check": [ "Most traffic uses the cheap path; failures add strong-model cost.", "Most traffic uses the cheap path; failures add strong-model cost." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/routing-1/", "date": "2026-10-26" }, { "id": "routing-2", "note": [ "Failed cheap calls are billed and fallback is assumed successful. Expected latency is serial and excludes queues; aggregate expectation is not p95 latency.", "失败的便宜调用仍计费，假设 fallback 成功。期望延迟按串行计算且忽略排队；总体期望不是 p95 延迟。" ], "check": [ "Strong-only routing provides a simple cost baseline.", "Strong-only routing provides a simple cost baseline." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/routing-2/", "date": "2026-10-26" }, { "id": "routing-3", "note": [ "Failed cheap calls are billed and fallback is assumed successful. Expected latency is serial and excludes queues; aggregate expectation is not p95 latency.", "失败的便宜调用仍计费，假设 fallback 成功。期望延迟按串行计算且忽略排队；总体期望不是 p95 延迟。" ], "check": [ "High fallback frequency can erase routing savings.", "High fallback frequency can erase routing savings." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/routing-3/", "date": "2026-10-26" }, { "id": "routing-4", "note": [ "Failed cheap calls are billed and fallback is assumed successful. Expected latency is serial and excludes queues; aggregate expectation is not p95 latency.", "失败的便宜调用仍计费，假设 fallback 成功。期望延迟按串行计算且忽略排队；总体期望不是 p95 延迟。" ], "check": [ "Every request pays both cost and latency before successful fallback.", "Every request pays both cost and latency before successful fallback." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/routing-4/", "date": "2026-10-26" }, { "id": "routing-5", "note": [ "Failed cheap calls are billed and fallback is assumed successful. Expected latency is serial and excludes queues; aggregate expectation is not p95 latency.", "失败的便宜调用仍计费，假设 fallback 成功。期望延迟按串行计算且忽略排队；总体期望不是 p95 延迟。" ], "check": [ "Expected cost is a weighted average without retries.", "Expected cost is a weighted average without retries." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/routing-5/", "date": "2026-10-26" }, { "id": "quantization-1", "note": [ "This is symmetric per-tensor rounding; metadata is excluded from ideal compression. All-zero tensors use a safe scale of 1. Do not interpret scalar reconstruction MSE as LLM quality.", "这是对称逐张量舍入；理想压缩率不含元数据。全零张量安全尺度取 1。标量重建 MSE 不是 LLM 质量。" ], "check": [ "Fewer representable values introduce rounding error.", "Fewer representable values introduce rounding error." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/quantization-1/", "date": "2026-10-26" }, { "id": "quantization-2", "note": [ "This is symmetric per-tensor rounding; metadata is excluded from ideal compression. All-zero tensors use a safe scale of 1. Do not interpret scalar reconstruction MSE as LLM quality.", "这是对称逐张量舍入；理想压缩率不含元数据。全零张量安全尺度取 1。标量重建 MSE 不是 LLM 质量。" ], "check": [ "More quantization levels usually reduce reconstruction error.", "More quantization levels usually reduce reconstruction error." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/quantization-2/", "date": "2026-10-26" }, { "id": "quantization-3", "note": [ "This is symmetric per-tensor rounding; metadata is excluded from ideal compression. All-zero tensors use a safe scale of 1. Do not interpret scalar reconstruction MSE as LLM quality.", "这是对称逐张量舍入；理想压缩率不含元数据。全零张量安全尺度取 1。标量重建 MSE 不是 LLM 质量。" ], "check": [ "Only three signed levels preserve little detail.", "Only three signed levels preserve little detail." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/quantization-3/", "date": "2026-10-26" }, { "id": "quantization-4", "note": [ "This is symmetric per-tensor rounding; metadata is excluded from ideal compression. All-zero tensors use a safe scale of 1. Do not interpret scalar reconstruction MSE as LLM quality.", "这是对称逐张量舍入；理想压缩率不含元数据。全零张量安全尺度取 1。标量重建 MSE 不是 LLM 质量。" ], "check": [ "An outlier enlarges the global step and collapses small values.", "An outlier enlarges the global step and collapses small values." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/quantization-4/", "date": "2026-10-26" }, { "id": "quantization-5", "note": [ "This is symmetric per-tensor rounding; metadata is excluded from ideal compression. All-zero tensors use a safe scale of 1. Do not interpret scalar reconstruction MSE as LLM quality.", "这是对称逐张量舍入；理想压缩率不含元数据。全零张量安全尺度取 1。标量重建 MSE 不是 LLM 质量。" ], "check": [ "The zero-range safeguard reconstructs all zeros exactly.", "The zero-range safeguard reconstructs all zeros exactly." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/quantization-5/", "date": "2026-10-26" }, { "id": "cache-1", "note": [ "TTL starts at insertion and is not refreshed on a hit. LRU order does refresh. Request keys share one synthetic namespace; production keys must include model, parameters and authorization scope.", "TTL 从插入开始且命中不刷新；LRU 顺序会刷新。实验键共享合成命名空间；生产键还须包含模型、参数和授权范围。" ], "check": [ "Repeated keys can avoid repeated model work.", "Repeated keys can avoid repeated model work." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/cache-1/", "date": "2026-10-26" }, { "id": "cache-2", "note": [ "TTL starts at insertion and is not refreshed on a hit. LRU order does refresh. Request keys share one synthetic namespace; production keys must include model, parameters and authorization scope.", "TTL 从插入开始且命中不刷新；LRU 顺序会刷新。实验键共享合成命名空间；生产键还须包含模型、参数和授权范围。" ], "check": [ "A tiny cache evicts useful entries before reuse.", "A tiny cache evicts useful entries before reuse." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/cache-2/", "date": "2026-10-26" }, { "id": "cache-3", "note": [ "TTL starts at insertion and is not refreshed on a hit. LRU order does refresh. Request keys share one synthetic namespace; production keys must include model, parameters and authorization scope.", "TTL 从插入开始且命中不刷新；LRU 顺序会刷新。实验键共享合成命名空间；生产键还须包含模型、参数和授权范围。" ], "check": [ "Entries expire at the next request with this interval.", "Entries expire at the next request with this interval." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/cache-3/", "date": "2026-10-26" }, { "id": "cache-4", "note": [ "TTL starts at insertion and is not refreshed on a hit. LRU order does refresh. Request keys share one synthetic namespace; production keys must include model, parameters and authorization scope.", "TTL 从插入开始且命中不刷新；LRU 顺序会刷新。实验键共享合成命名空间；生产键还须包含模型、参数和授权范围。" ], "check": [ "Temporal locality improves the hit rate.", "Temporal locality improves the hit rate." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/cache-4/", "date": "2026-10-26" }, { "id": "cache-5", "note": [ "TTL starts at insertion and is not refreshed on a hit. LRU order does refresh. Request keys share one synthetic namespace; production keys must include model, parameters and authorization scope.", "TTL 从插入开始且命中不刷新；LRU 顺序会刷新。实验键共享合成命名空间；生产键还须包含模型、参数和授权范围。" ], "check": [ "Unique keys provide no exact-match cache savings.", "Unique keys provide no exact-match cache savings." ], "files": [ "academy-core.mjs" ], "scope": [ "Engine formulas, scenario presets and numerical safeguards reviewed. Seeded synthetic inputs are not a production benchmark.", "已复核引擎公式、场景预设及数值保护。固定种子的合成输入不是生产基准。" ], "path": "learn/cache-5/", "date": "2026-10-26" }, { "id": "gradient-debugger", "note": [ "Synthetic linear regression, not a deep-network profiler. Training stops if loss exceeds 10¹². Finite-difference checks are local and do not prove every gradient is correct.", "这是合成数据线性回归，不是深度网络性能分析器。损失超过 10¹² 即停止。有限差分只是局部检查，不能证明所有位置的梯度都正确。" ], "check": [ "Run the baseline; enable the reduction bug; then disable it and set input scale to 20. Find a learning rate that restores stable training. Compare exports with one variable changed.", "先运行基线，再启用归约错误；之后关闭错误并把输入尺度设为 20。寻找恢复稳定训练的学习率，每次只改变一个变量并对比导出记录。" ], "files": [ "phd-core.mjs", "academy-core.mjs", "research-core.mjs" ], "scope": [ "Objective, derivatives, worked examples and implementation reviewed. A passing finite-difference or manifest check is not a paper reproduction certificate.", "已复核目标、导数、算例及实现。通过有限差分或清单检查不代表完成论文复现。" ], "path": "research/gradient-debugger/", "date": "2026-10-26" }, { "id": "information-theory", "note": [ "This compares categorical distributions, not empirical model calibration. Calibration also needs predictions paired with observed outcomes; entropy alone is not correctness.", "这里比较分类分布，不进行模型校准评估。校准还需要预测与真实结果配对；熵不等于正确率。" ], "check": [ "Set P=Q, then use P=1,0 and Q=0,1. Explain the infinite KL and finite Jensen–Shannon divergence without introducing an arbitrary epsilon.", "先设置 P=Q，再用 P=1,0 和 Q=0,1。解释为什么 KL 无穷大而 JS 有限，不要用任意 epsilon 掩盖问题。" ], "files": [ "phd-core.mjs", "academy-core.mjs", "research-core.mjs" ], "scope": [ "Objective, derivatives, worked examples and implementation reviewed. A passing finite-difference or manifest check is not a paper reproduction certificate.", "已复核目标、导数、算例及实现。通过有限差分或清单检查不代表完成论文复现。" ], "path": "research/information-theory/", "date": "2026-10-26" }, { "id": "lora-rank", "note": [ "Actual local matrix optimization, not fine-tuning a language model. Frozen base W=0 and α/r=1. Changing the parameter-calculator dimension does not resize the 8×8 experiment. Equal steps are not equal FLOPs across ranks.", "这是真实的本地矩阵优化，不是语言模型微调。固定基座 W=0，α/r=1。修改参数量计算器的维度不会改变 8×8 实验。不同秩的相同步数不等于相同 FLOPs。" ], "check": [ "Compare rank 1,2,4 with identical seed and steps. Then initialize both factors to zero. Separate the theoretical rank limit from failure to reach that limit during optimization.", "使用相同种子和步数比较 rank 1、2、4，再将两个因子均初始化为零。区分秩带来的理论下界与优化尚未达到下界。" ], "files": [ "phd-core.mjs", "academy-core.mjs", "research-core.mjs" ], "scope": [ "Objective, derivatives, worked examples and implementation reviewed. A passing finite-difference or manifest check is not a paper reproduction certificate.", "已复核目标、导数、算例及实现。通过有限差分或清单检查不代表完成论文复现。" ], "path": "research/lora-rank/", "date": "2026-10-26" }, { "id": "experiment-design", "note": [ "95% paired percentile bootstrap, 2000 resamples, at most 200 rows. Small samples and constant deltas can be misleading. No multiple-testing correction, power calculation or automatic causal inference.", "95% 配对百分位 bootstrap，2000 次重采样，最多 200 行。小样本和恒定差值可能误导。本工具不做多重检验校正、功效计算或自动因果推断。" ], "check": [ "Analyze the sample, reverse metric direction, and duplicate a seed ID to observe validation. Then compare a noisy run set with the same mean improvement. Explain why mean alone is insufficient.", "分析示例，反转指标方向，并重复种子 ID 观察校验。再比较均值提升相同但噪声更大的数据，解释为什么不能只报告均值。" ], "files": [ "phd-core.mjs", "academy-core.mjs", "research-core.mjs" ], "scope": [ "Objective, derivatives, worked examples and implementation reviewed. A passing finite-difference or manifest check is not a paper reproduction certificate.", "已复核目标、导数、算例及实现。通过有限差分或清单检查不代表完成论文复现。" ], "path": "research/experiment-design/", "date": "2026-10-26" }, { "id": "contrastive-learning", "note": [ "Exact single-anchor loss and score gradients with supplied similarities; no encoder is trained. The tool cannot identify false negatives automatically. Compare retrieval quality using a separate dataset.", "这里精确计算单锚点损失与分数梯度，不训练编码器，也不能自动识别假负样本。检索质量需要独立数据集验证。" ], "check": [ "Treat index 1 as a known false negative. Run with all candidates, then exclude index 1 and compare its gradient and the positive probability. Sweep temperature 0.05, 0.2 and 1.", "假设索引 1 是已知假负样本。先使用所有候选，再排除索引 1，对比它的梯度和正样本概率。依次测试温度 0.05、0.2 和 1。" ], "files": [ "phd-core.mjs", "academy-core.mjs", "research-core.mjs" ], "scope": [ "Objective, derivatives, worked examples and implementation reviewed. A passing finite-difference or manifest check is not a paper reproduction certificate.", "已复核目标、导数、算例及实现。通过有限差分或清单检查不代表完成论文复现。" ], "path": "research/contrastive-learning/", "date": "2026-10-26" }, { "id": "paper-reproduction", "note": [ "No paper fetch, cloud training or provenance verification. The included runs are synthetic. Nested configuration differences are counted by top-level key; several causal changes can hide inside one key. Manual review remains required.", "不抓取论文、不运行云端训练、不验证来源。内置数据为合成示例。嵌套配置按顶层键统计变化，一个键内可能隐藏多项因果变化，仍需人工审查。" ], "check": [ "Replace the demo metadata with actual identifiers, paste paired seed results, then introduce a second configuration change or budget mismatch. Export the report and explain which concerns remain manual.", "用真实标识替换示例元数据，粘贴配对种子结果，再增加第二项配置变化或制造预算不一致。导出报告，解释哪些问题仍需人工审查。" ], "files": [ "phd-core.mjs", "academy-core.mjs", "research-core.mjs" ], "scope": [ "Objective, derivatives, worked examples and implementation reviewed. A passing finite-difference or manifest check is not a paper reproduction certificate.", "已复核目标、导数、算例及实现。通过有限差分或清单检查不代表完成论文复现。" ], "path": "research/paper-reproduction/", "date": "2026-10-26" }, { "id": "prompt-diff", "note": [ "Line edits and token deltas are different quantities. Whitespace affects tokenization; a line edit is not semantic equivalence.", "行编辑数与 token 差值不同。空白影响分词；行编辑不是语义等价判断。" ], "check": [ "Line edits and token deltas are different quantities. Whitespace affects tokenization; a line edit is not semantic equivalence.", "行编辑数与 token 差值不同。空白影响分词；行编辑不是语义等价判断。" ], "files": [ "research-core.mjs" ], "dependency": true, "scope": [ "Reviewed counting units, statistical assumptions and numerical examples. Provider accounting and actual GPU allocation still require external measurement.", "已复核计数单位、统计假设及数值例子；提供商计费和真实 GPU 分配仍须实测。" ], "path": "tools/prompt-diff/", "date": "2026-10-26" }, { "id": "jsonl-inspector", "note": [ "Exact normalized text overlap is a leakage signal, not proof that all near duplicates are detected. Token counts exclude provider chat-template overhead.", "规范化文本精确重叠是泄漏信号，不证明检出了所有近重复；token 计数不含提供商聊天模板开销。" ], "check": [ "Exact normalized text overlap is a leakage signal, not proof that all near duplicates are detected. Token counts exclude provider chat-template overhead.", "规范化文本精确重叠是泄漏信号，不证明检出了所有近重复；token 计数不含提供商聊天模板开销。" ], "files": [ "research-core.mjs" ], "dependency": true, "scope": [ "Reviewed counting units, statistical assumptions and numerical examples. Provider accounting and actual GPU allocation still require external measurement.", "已复核计数单位、统计假设及数值例子；提供商计费和真实 GPU 分配仍须实测。" ], "path": "tools/jsonl-inspector/", "date": "2026-10-26" }, { "id": "rag-chunker", "note": [ "Chunk size and overlap are Unicode code points, not tokens or UTF-8 bytes. Reconstruct the input without duplicating overlaps; then separately count tokens.", "块长和重叠单位为 Unicode 码点，不是 token 或 UTF-8 字节。去掉重叠后应重建原文，再独立计 token。" ], "check": [ "Chunk size and overlap are Unicode code points, not tokens or UTF-8 bytes. Reconstruct the input without duplicating overlaps; then separately count tokens.", "块长和重叠单位为 Unicode 码点，不是 token 或 UTF-8 字节。去掉重叠后应重建原文，再独立计 token。" ], "files": [ "research-core.mjs" ], "dependency": true, "scope": [ "Reviewed counting units, statistical assumptions and numerical examples. Provider accounting and actual GPU allocation still require external measurement.", "已复核计数单位、统计假设及数值例子；提供商计费和真实 GPU 分配仍须实测。" ], "path": "tools/rag-chunker/", "date": "2026-10-26" }, { "id": "eval-compare", "note": [ "Resample paired independent seed/run means, not correlated rows from one run. Percentile bootstrap is not a causal claim or a p-value; a degenerate fixture can produce a zero-width interval.", "重采样成对独立种子/运行均值，不是同一运行相关行。百分位 bootstrap 不是因果结论或 p 值；退化样例会产生零宽区间。" ], "check": [ "Resample paired independent seed/run means, not correlated rows from one run. Percentile bootstrap is not a causal claim or a p-value; a degenerate fixture can produce a zero-width interval.", "重采样成对独立种子/运行均值，不是同一运行相关行。百分位 bootstrap 不是因果结论或 p 值；退化样例会产生零宽区间。" ], "files": [ "research-core.mjs" ], "dependency": true, "scope": [ "Reviewed counting units, statistical assumptions and numerical examples. Provider accounting and actual GPU allocation still require external measurement.", "已复核计数单位、统计假设及数值例子；提供商计费和真实 GPU 分配仍须实测。" ], "path": "tools/eval-compare/", "date": "2026-10-26" }, { "id": "vram-calculator", "note": [ "The 18-byte training baseline is 2 working-weight + 4 master-weight + 4 gradient + 8 Adam-moment bytes per parameter. This is one dtype layout, not a universal framework constant or measured peak.", "每参数 18 字节基线为 2 字节工作权重+4 主权重+4 梯度+8 Adam 动量。这是一种 dtype 布局，不是通用框架常量或实测峰值。" ], "check": [ "The 18-byte training baseline is 2 working-weight + 4 master-weight + 4 gradient + 8 Adam-moment bytes per parameter. This is one dtype layout, not a universal framework constant or measured peak.", "每参数 18 字节基线为 2 字节工作权重+4 主权重+4 梯度+8 Adam 动量。这是一种 dtype 布局，不是通用框架常量或实测峰值。" ], "files": [ "research-core.mjs" ], "dependency": true, "scope": [ "Reviewed counting units, statistical assumptions and numerical examples. Provider accounting and actual GPU allocation still require external measurement.", "已复核计数单位、统计假设及数值例子；提供商计费和真实 GPU 分配仍须实测。" ], "path": "tools/vram-calculator/", "date": "2026-10-26" }, { "id": "sampling-lab", "note": [ "Filtering order is temperature, top-k, then top-p and renormalization. Entropy is in bits. Temperature zero is explicitly greedy; stochastic draws depend on the seed.", "过滤顺序为温度、top-k、top-p，再归一化。熵单位 bits；零温度明确采用贪心，随机采样依赖种子。" ], "check": [ "Filtering order is temperature, top-k, then top-p and renormalization. Entropy is in bits. Temperature zero is explicitly greedy; stochastic draws depend on the seed.", "过滤顺序为温度、top-k、top-p，再归一化。熵单位 bits；零温度明确采用贪心，随机采样依赖种子。" ], "files": [ "research-core.mjs" ], "dependency": true, "scope": [ "Reviewed counting units, statistical assumptions and numerical examples. Provider accounting and actual GPU allocation still require external measurement.", "已复核计数单位、统计假设及数值例子；提供商计费和真实 GPU 分配仍须实测。" ], "path": "tools/sampling-lab/", "date": "2026-10-26" }, { "id": "speculative-decoding", "note": [ "Expected emissions 1+α+…+α^γ assume independent identical acceptance and a fixed verifier latency. This latency simulator is distinct from exact rejection-sampling correctness.", "期望产出 1+α+…+α^γ 假设独立等接受率及固定验证延迟；此延迟模拟不同于精确拒绝采样正确性。" ], "check": [ "Expected emissions 1+α+…+α^γ assume independent identical acceptance and a fixed verifier latency. This latency simulator is distinct from exact rejection-sampling correctness.", "期望产出 1+α+…+α^γ 假设独立等接受率及固定验证延迟；此延迟模拟不同于精确拒绝采样正确性。" ], "files": [ "research-core.mjs" ], "dependency": true, "scope": [ "Reviewed counting units, statistical assumptions and numerical examples. Provider accounting and actual GPU allocation still require external measurement.", "已复核计数单位、统计假设及数值例子；提供商计费和真实 GPU 分配仍须实测。" ], "path": "tools/speculative-decoding/", "date": "2026-10-26" }, { "id": "tokens-and-tokenizers", "note": [ "Round-trip decode(encode(text)) equals the original text. Individual tokens may split UTF-8 sequences; do not require each token to decode independently.", "整段 decode(encode(text)) 应恢复原文。单 token 可拆开 UTF-8 序列，不能要求各 token 独立解码无损。" ], "check": [ "Round-trip decode(encode(text)) equals the original text. Individual tokens may split UTF-8 sequences; do not require each token to decode independently.", "整段 decode(encode(text)) 应恢复原文。单 token 可拆开 UTF-8 序列，不能要求各 token 独立解码无损。" ], "files": [], "dependency": true, "scope": [ "Shared code, English mathematical explanation and key multilingual numeric passages reviewed. This is not a full translation proofread.", "已复核共享代码、英文数学解释及多语言关键数值段落；不代表完成逐字翻译校对。" ], "path": "tutorials/tokens-and-tokenizers/", "date": "2026-10-26" }, { "id": "multilingual-tokenization", "note": [ "Hold encoding and text fixed. Record UTF-8 bytes, code points and tokens separately; the English 4-characters-per-token rule is not a multilingual formula.", "固定编码与输入，分别记录 UTF-8 字节、码点和 token；英文每 token 约 4 字符不是多语言公式。" ], "check": [ "Hold encoding and text fixed. Record UTF-8 bytes, code points and tokens separately; the English 4-characters-per-token rule is not a multilingual formula.", "固定编码与输入，分别记录 UTF-8 字节、码点和 token；英文每 token 约 4 字符不是多语言公式。" ], "files": [], "dependency": true, "scope": [ "Shared code, English mathematical explanation and key multilingual numeric passages reviewed. This is not a full translation proofread.", "已复核共享代码、英文数学解释及多语言关键数值段落；不代表完成逐字翻译校对。" ], "path": "tutorials/multilingual-tokenization/", "date": "2026-10-26" }, { "id": "context-budget-rag", "note": [ "8096 of 8192 leaves 96 tokens. Adding 500 overflows by 404; to also preserve a 256-token reserve, remove 660. Provider template overhead must be measured separately.", "8192 中已分配 8096，剩 96；再加 500 超出 404。还要保留 256 安全余量需删 660；提供商模板开销须另计。" ], "check": [ "8096 of 8192 leaves 96 tokens. Adding 500 overflows by 404; to also preserve a 256-token reserve, remove 660. Provider template overhead must be measured separately.", "8192 中已分配 8096，剩 96；再加 500 超出 404。还要保留 256 安全余量需删 660；提供商模板开销须另计。" ], "files": [], "dependency": false, "scope": [ "Shared code, English mathematical explanation and key multilingual numeric passages reviewed. This is not a full translation proofread.", "已复核共享代码、英文数学解释及多语言关键数值段落；不代表完成逐字翻译校对。" ], "path": "tutorials/context-budget-rag/", "date": "2026-10-26" }, { "id": "base64-ai-payload-inspector", "note": [ "Base64 length is 4 ceil(n/3): byte lengths 0,1,2,3 produce 0,4,4,4 characters. Parent JSON size already includes children; summing both double-counts. Bytes are not model tokens.", "Base64 长度 4 ceil(n/3)：0、1、2、3 字节对应 0、4、4、4 字符。父 JSON 大小包含子节点，再累加会重复；字节不是模型 token。" ], "check": [ "Base64 length is 4 ceil(n/3): byte lengths 0,1,2,3 produce 0,4,4,4 characters. Parent JSON size already includes children; summing both double-counts. Bytes are not model tokens.", "Base64 长度 4 ceil(n/3)：0、1、2、3 字节对应 0、4、4、4 字符。父 JSON 大小包含子节点，再累加会重复；字节不是模型 token。" ], "files": [], "dependency": false, "scope": [ "Shared code, English mathematical explanation and key multilingual numeric passages reviewed. This is not a full translation proofread.", "已复核共享代码、英文数学解释及多语言关键数值段落；不代表完成逐字翻译校对。" ], "path": "tutorials/base64-ai-payload-inspector/", "date": "2026-10-26" }, { "id": "robot-recovery", "note": [ "Seven deterministic recovery transition checks; scripted risk inputs, not a trained VLA or physical safety controller.", "七项确定性恢复状态检查；脚本风险输入，不代表训练后的 VLA 或真实安全控制器。" ], "check": [ "Seven deterministic recovery transition checks; scripted risk inputs, not a trained VLA or physical safety controller.", "七项确定性恢复状态检查；脚本风险输入，不代表训练后的 VLA 或真实安全控制器。" ], "files": [ "robot-recovery-core.mjs" ], "copySource": "examples/robot-recovery.mjs", "scope": [ "Implementation, invariants and declared evidence boundary reviewed; existing acceptance tests remain part of the suite.", "已复核实现、不变性及声明的证据边界；保留原有验收测试。" ], "path": "frontier/robot-recovery/", "date": "2026-10-26" }, { "id": "adaptive-agent-evaluation", "note": [ "Population normalization, zero variance and matched-budget selection checks. Synthetic fixture, not a paper reproduction.", "总体方差归一化、零方差与等预算选择检查；合成教学样例，并非论文复现。" ], "check": [ "Population normalization, zero variance and matched-budget selection checks. Synthetic fixture, not a paper reproduction.", "总体方差归一化、零方差与等预算选择检查；合成教学样例，并非论文复现。" ], "files": [ "adaptive-agent-evaluation-core.mjs" ], "copySource": "examples/adaptive-harness.mjs", "scope": [ "Implementation, invariants and declared evidence boundary reviewed; existing acceptance tests remain part of the suite.", "已复核实现、不变性及声明的证据边界；保留原有验收测试。" ], "path": "frontier/adaptive-agent-evaluation/", "date": "2026-10-26" }, { "id": "agent-harness-engineering", "note": [ "Seven deterministic execution-gate checks; simulated effects only. No model capability or exactly-once external effect claim.", "七项确定性执行门检查；仅模拟副作用，不代表模型能力或外部效果恰好一次保证。" ], "check": [ "Seven deterministic execution-gate checks; simulated effects only. No model capability or exactly-once external effect claim.", "七项确定性执行门检查；仅模拟副作用，不代表模型能力或外部效果恰好一次保证。" ], "files": [ "agent-harness-engineering-core.mjs" ], "copySource": "examples/agent-harness.mjs", "scope": [ "Implementation, invariants and declared evidence boundary reviewed; existing acceptance tests remain part of the suite.", "已复核实现、不变性及声明的证据边界；保留原有验收测试。" ], "path": "frontier/agent-harness-engineering/", "date": "2026-10-26" }, { "id": "application-document-evidence", "note": [ "Lexical match score divides matched unique query terms by unique query-term count. Empty or unmatched queries abstain. Evidence quotation is not semantic entailment.", "词法分数为匹配的唯一查询词数/唯一查询词数。空查询或无匹配则拒答。引用证据不等于语义蕴含。" ], "check": [ "Lexical match score divides matched unique query terms by unique query-term count. Empty or unmatched queries abstain. Evidence quotation is not semantic entailment.", "词法分数为匹配的唯一查询词数/唯一查询词数。空查询或无匹配则拒答。引用证据不等于语义蕴含。" ], "files": [ "application-document-evidence-core.mjs" ], "copySource": "examples/document-evidence.mjs", "scope": [ "Implementation, invariants and declared evidence boundary reviewed; existing acceptance tests remain part of the suite.", "已复核实现、不变性及声明的证据边界；保留原有验收测试。" ], "path": "applications/document-evidence/", "date": "2026-10-26" }, { "id": "application-validated-extraction", "note": [ "Shape validation and exact evidence checks are separate. Decimal display uses two digits but is not exact monetary arithmetic. Substrings cannot resolve conflicting invoice totals.", "结构校验与精确证据校验分开。小数显示保留两位，不是精确货币算术。子串匹配无法解决发票总额冲突。" ], "check": [ "Shape validation and exact evidence checks are separate. Decimal display uses two digits but is not exact monetary arithmetic. Substrings cannot resolve conflicting invoice totals.", "结构校验与精确证据校验分开。小数显示保留两位，不是精确货币算术。子串匹配无法解决发票总额冲突。" ], "files": [ "application-validated-extraction-core.mjs" ], "copySource": "examples/validated-extraction.mjs", "scope": [ "Implementation, invariants and declared evidence boundary reviewed; existing acceptance tests remain part of the suite.", "已复核实现、不变性及声明的证据边界；保留原有验收测试。" ], "path": "applications/validated-extraction/", "date": "2026-10-26" }, { "id": "application-bounded-tools", "note": [ "Only executed calls consume budget. Replays require both matching call ID and matching arguments. Allowlist failure blocks that call; it does not terminate the supplied plan.", "只有实际执行消耗预算。重放要求调用 ID 和参数一致。白名单失败阻止该调用，不自动终止整个给定计划。" ], "check": [ "Only executed calls consume budget. Replays require both matching call ID and matching arguments. Allowlist failure blocks that call; it does not terminate the supplied plan.", "只有实际执行消耗预算。重放要求调用 ID 和参数一致。白名单失败阻止该调用，不自动终止整个给定计划。" ], "files": [ "application-bounded-tools-core.mjs" ], "copySource": "examples/bounded-tools.mjs", "scope": [ "Implementation, invariants and declared evidence boundary reviewed; existing acceptance tests remain part of the suite.", "已复核实现、不变性及声明的证据边界；保留原有验收测试。" ], "path": "applications/bounded-tools/", "date": "2026-10-26" }, { "id": "project-attention", "note": [ "Fixed self-attention with finite Q/K/V, nonempty consistent dimensions, row-wise masking and no dropout. Prefix caching must equal full causal output; future perturbations must leave earlier outputs unchanged.", "固定自注意力，Q/K/V 有限且维度一致非空，逐行遮罩，无 dropout。前缀缓存须与完整因果输出一致；未来扰动不得改变较早输出。" ], "check": [ "Fixed self-attention with finite Q/K/V, nonempty consistent dimensions, row-wise masking and no dropout. Prefix caching must equal full causal output; future perturbations must leave earlier outputs unchanged.", "固定自注意力，Q/K/V 有限且维度一致非空，逐行遮罩，无 dropout。前缀缓存须与完整因果输出一致；未来扰动不得改变较早输出。" ], "files": [ "project-attention-core.mjs" ], "copySource": "projects/attention.mjs", "scope": [ "Implementation, invariants and declared evidence boundary reviewed; existing acceptance tests remain part of the suite.", "已复核实现、不变性及声明的证据边界；保留原有验收测试。" ], "path": "projects/attention/", "date": "2026-10-26" }, { "id": "project-retrieval", "note": [ "Five development queries: baseline 2/5 and tag-assisted 5/5 accuracy. Tags were authored for these concepts. This is an engineering fixture, not held-out generalization or trained embedding retrieval.", "五条开发查询：基线 2/5，标签辅助 5/5。标签针对这些概念编写。这是工程样例，不是留出泛化或训练 embedding 检索。" ], "check": [ "Five development queries: baseline 2/5 and tag-assisted 5/5 accuracy. Tags were authored for these concepts. This is an engineering fixture, not held-out generalization or trained embedding retrieval.", "五条开发查询：基线 2/5，标签辅助 5/5。标签针对这些概念编写。这是工程样例，不是留出泛化或训练 embedding 检索。" ], "files": [ "project-retrieval-core.mjs" ], "copySource": "projects/retrieval.mjs", "scope": [ "Implementation, invariants and declared evidence boundary reviewed; existing acceptance tests remain part of the suite.", "已复核实现、不变性及声明的证据边界；保留原有验收测试。" ], "path": "projects/retrieval/", "date": "2026-10-26" }, { "id": "project-control", "note": [ "Explicit Euler uses old velocity for the position update. Force saturation is in newtons; delay is an integer number of steps. No contact dynamics, collisions or real hardware guarantees are modeled.", "显式 Euler 用旧速度更新位置。力饱和单位牛顿，延迟为整数步数。不建模接触动力学、碰撞，也不保证真实硬件表现。" ], "check": [ "Explicit Euler uses old velocity for the position update. Force saturation is in newtons; delay is an integer number of steps. No contact dynamics, collisions or real hardware guarantees are modeled.", "显式 Euler 用旧速度更新位置。力饱和单位牛顿，延迟为整数步数。不建模接触动力学、碰撞，也不保证真实硬件表现。" ], "files": [ "project-control-core.mjs" ], "copySource": "projects/control.mjs", "scope": [ "Implementation, invariants and declared evidence boundary reviewed; existing acceptance tests remain part of the suite.", "已复核实现、不变性及声明的证据边界；保留原有验收测试。" ], "path": "projects/control/", "date": "2026-10-26" } ] }