From a9902b42d078f0f631a6b12e3c21272eb8b3ca24 Mon Sep 17 00:00:00 2001 From: Frank14f <1515444314@qq.com> Date: Sat, 14 Nov 2026 17:15:58 +0800 Subject: [PATCH] Document S2 execution differences against unified v0.1 and preserve stage review policy --- AGENTS.md | 8 + docs/collaboration-stage-review-policy.md | 19 ++ docs/s2-execution-diff-v0-2026-11-14.md | 169 ++++++++++++++++++ docs/sharing-authorization-2026-11-14.md | 11 ++ .../s2-design/claude-data-contract-v0.md | 74 ++++++++ .../2026-11-14/s2-design/claude-h1-v01.md | 36 ++++ .../s2-design/claude-research-v2.md | 137 ++++++++++++++ .../s2-design/claude-research-v21.md | 45 +++++ .../s2-design/claude-s2-handoff-source.txt | 67 +++++++ .../s2-design/claude-s2-v0-d8aa0f42.md | 65 +++++++ .../s2-design/claude-s2-v01-fd1db63e.md | 65 +++++++ .../s2-design/claude-stage-review-3ae89805.md | 61 +++++++ .../s2-design/document-manifest.json | 100 +++++++++++ .../2026-11-14/s2-design/moldastra-note13.md | 18 ++ .../2026-11-14/s2-design/moldastra-note14.md | 18 ++ .../2026-11-14/s2-design/moldastra-note15.md | 18 ++ .../s1-review-and-topology-steering.txt | 45 +++++ ...review-user-rule-and-target-correction.txt | 82 +++++++++ .../v01-grouping-and-stop-corrections.txt | 82 +++++++++ 19 files changed, 1120 insertions(+) create mode 100644 AGENTS.md create mode 100644 docs/collaboration-stage-review-policy.md create mode 100644 docs/s2-execution-diff-v0-2026-11-14.md create mode 100644 docs/sharing-authorization-2026-11-14.md create mode 100644 references/2026-11-14/s2-design/claude-data-contract-v0.md create mode 100644 references/2026-11-14/s2-design/claude-h1-v01.md create mode 100644 references/2026-11-14/s2-design/claude-research-v2.md create mode 100644 references/2026-11-14/s2-design/claude-research-v21.md create mode 100644 references/2026-11-14/s2-design/claude-s2-handoff-source.txt create mode 100644 references/2026-11-14/s2-design/claude-s2-v0-d8aa0f42.md create mode 100644 references/2026-11-14/s2-design/claude-s2-v01-fd1db63e.md create mode 100644 references/2026-11-14/s2-design/claude-stage-review-3ae89805.md create mode 100644 references/2026-11-14/s2-design/document-manifest.json create mode 100644 references/2026-11-14/s2-design/moldastra-note13.md create mode 100644 references/2026-11-14/s2-design/moldastra-note14.md create mode 100644 references/2026-11-14/s2-design/moldastra-note15.md create mode 100644 references/2026-11-14/s2-design/s1-review-and-topology-steering.txt create mode 100644 references/2026-11-14/s2-design/stage-review-user-rule-and-target-correction.txt create mode 100644 references/2026-11-14/s2-design/v01-grouping-and-stop-corrections.txt diff --git a/AGENTS.md b/AGENTS.md new file mode 100644 index 000000000..72971519b --- /dev/null +++ b/AGENTS.md @@ -0,0 +1,8 @@ +# MeltMind project collaboration + +- Before moving on from a major result or entering a new stage, follow `docs/collaboration-stage-review-policy.md`: consolidate evidence across the team, revisit the user's research goal and prior work, discuss disagreements, and record the next-stage decision. +- Continue the existing shared design and its fixed commits. Check for an existing draft or running task before starting another. Claude integrates and reviews; MeltAstra supplies execution evidence; MoldAstra supplies mathematical/grouping boundaries; Grok performs only needed targeted checks. +- Keep topology/local-type partitioning and true connectivity as independent open hypotheses. A failed input-coverage mechanism is not a failure of all partition or topology routes. +- Design/document review and information sharing do not authorize new scientific computation or training. Carry out the scope already authorized; record explicit scope and budgets before expanding it. +- The user has authorized project information sharing in their controlled Drive, Git and Slack spaces; see `docs/sharing-authorization-2026-11-14.md`. Do not ask for the same sharing authorization again. Git identity discussion is not a priority. +- Preserve original data, failures, review limitations and immutable evidence. Do not label proposals as executed or reviewed results. diff --git a/docs/collaboration-stage-review-policy.md b/docs/collaboration-stage-review-policy.md new file mode 100644 index 000000000..ba6d87129 --- /dev/null +++ b/docs/collaboration-stage-review-policy.md @@ -0,0 +1,19 @@ +# MeltMind 阶段回顾与推进规则 + +来源:Yanqi 经 TRON 于 2026-11-14 17:02 在[原 Slack 串](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794646921021699)同步的明确要求;本会话亦收到接续指令。本文件保存持续项目工作习惯。 + +**每逢重大结果,或准备进入下一阶段,先全体汇总,回顾研究初衷和前人工作,深入讨论后再推进。** 文稿整理、资料分享、评审通过与实际计算/训练放行分别记录。 + +阶段回顾至少包含: + +1. 初衷:面向新件的实用注塑填充代理;分类、切分与几何组织服务于泛化和样本效率,不先验认定收益。 +2. 本阶段证据:支持、不支持、根本未检验的命题;失败、资格/覆盖、对照问题及审查实际读到的范围。 +3. 前人工作:已有方法已做了什么,本项目尚缺的强基线与验证;区分已读原文、团队既有记录、厂商声称和待核实条目,不用不同数据/任务的数字拼成优势。 +4. 候选路线:保留仍未检验的独立问题,尤其 O1 的类型分块、真实连通和接口;不把某个覆盖指标失败扩大为拓扑路线失败。 +5. 分歧与下一步:最小可辨识问题、输入/标签/划分、主指标和停止条件、资源预算、待裁量项、负责人及明确的进入决定。 + +Claude 统筹整合与主审;MeltAstra 提供执行、数据和预算证据;MoldAstra 提供数学、类型与分组边界;Grok 只对现有材料中确有必要的关键遗漏定点复核,不重复全面检索或整篇审查。全体参与不意味着四人重复相同工作。 + +各方先查已有文稿和固定提交,在共同基线上补充;新意见保留来源和决定记录。Git 身份问题不再作为讨论重点。项目 Drive、Git、Slack 信息共享沿用用户已经给出的授权,不重复询问;分享授权不等同于新计算或训练授权。 + +阶段包须给出固定版本、证据链接、分歧处理及尚未满足的进入条件。没有明确放行时,继续可独立完成的设计/讨论,保持执行未启动。S1→S2 当前以 Claude `d8aa0f42` 的设计、`fd1db63e` 的 v0.1 修订和 `3ae89805` 的阶段回顾为共同基线,不另建第二套方案。 diff --git a/docs/s2-execution-diff-v0-2026-11-14.md b/docs/s2-execution-diff-v0-2026-11-14.md new file mode 100644 index 000000000..444c52446 --- /dev/null +++ b/docs/s2-execution-diff-v0-2026-11-14.md @@ -0,0 +1,169 @@ +# 对 Claude S2 v0+v0.1 的执行差分(供 v0.2 整合) + +**基线:** [d8aa0f420647701709e03629326dce8339db739a / s2-design-v0](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/d8aa0f420647701709e03629326dce8339db739a/docs/claude/s2-design-v0-2026-11-14.md)。同时接续 [3ae89805 阶段回顾](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/3ae898056734bdaaeda0bb77f3e50f324e5a6d8a/docs/claude/stage-review-S1-to-S2-2026-11-14.md)。修订基线为 [fd1db63ee1c4fb4415a36ed95a2eedfb73a5c85c / S2 v0.1](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/fd1db63ee1c4fb4415a36ed95a2eedfb73a5c85c/docs/claude/s2-design-v0.1-2026-11-14.md)。作者 MeltAstra,供 Claude 合入 v0.2;本轮只有文档和资料回读,无新增数据计算、拟合、训练或 S1 重跑。 + +## 0. 保留的统一设计与本增补范围 + +保留 v0 的研究问题和 **R2 / N / E** 三臂、N/E 同主干同 g12/g23、O2 仅在抽样查询节点进入 decoder、最终来源组件上的 **5-fold CV**、训练池 **25/50/100%** 从头训练及三个种子。R2 是 CPU 模型;“同主干”仅指 N/E。 + +本稿只补近重复/分组执行流程、R2 特征与成本、pilot 主干候选、统计/覆盖和预算进入条件。已有较宽本地草稿已标为被取代并保留,不作为另一方案发布。O1 及内部 CAE 对照按统一设计保持开放,未加入当前运行矩阵。 + +## 1. 对 v0 各节的差分登记 + +| ID | 对应统一稿 | v0.1 已吸纳 | 尚待裁量的执行项 | +|---|---|---|---| +| D1 | v0 §3 / v0.1 B | 秒单位目标、训练池限定变换、禁用测试 Tmax 与 LOG 工艺输出 | 每外折/预算的尺度拟合范围(本文 §2) | +| D2 | v0 §3 / v0.1 C | 同来源同侧、几何证据边连通分量、249键非独立家族 | 定标/审边与模板关系规则、内层验证20%建议、既有研究暴露标记 | +| D3 | v0 §3 / v0.1 B | d>0 的S1质量PPS、报告覆盖、边界域排除 | 逐项资格、重复次数权重、零分母和有效域、固定查询及辅助指标支持域 | +| D4 | v0 §2 / v0.1 D | Gre24配方在本项目分组上执行,公开数字不作外部门槛 | 精确配方对应及本稿R2-G候选范围、权重/HPO与首次准备成本;不能将近似候选自动称为复现 | +| D5 | v0 §2 | 全节点encoder、同主干同g、E四项边读数进入query decoder | 源码/许可、容量与输入列参数差、V100 pilot检查 | +| D6 | v0 §4 / v0.1 F | 冻结m、种子与切片 | 主预算、相对减少估计量、匹配seed-block/组件区间与多比较 | +| D7 | v0 §5 / v0.1 F | 先pilot及明确go | 90神经+45CPU分类记账、有限准备/pilot资源表、不得用情景外推作实测预算 | +| D8 | v0.1 D/E | 新增停止/薄区/连通诊断条目 | D的数学修正由MoldAstra负责,本文只记录状态;E的全图预测/O2支持与费用尚未解决(§4.1) | + +v0.1 A 对研究范围的限缩已吸纳:S2只检验四项边方向读数增量,不能代替类型组织或O1检验。v0.1 D 的“1%时间分辨率→2%不可辨识”已被Claude在[17:12:56](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794647576007729)降为待证据规则;[MoldAstra note15](https://drive.google.com/file/d/18X07uNtM1j697MCztXryzmITCcud3_T4/view)已提交替换建议:有独立工程依据才用预算停止线,触发为“E−N未检验”,否则取消强制停止。此处不重复其证明,也不将尚待Claude写入的建议冒称固定稿已修正。 + +## 2. D1:已采纳的目标泄漏修正 + +[Claude 的修正原文](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794647015752699):主目标 Fill_time 使用秒,不以测试件真实 Tmax 归一。可选训练变换 `log(1+T/τ)` 的 τ 仅从**当前外折、当前预算训练池**拟合,预测反变换到秒后评价。较大训练池标签不能为较小预算提供标度。真实 T/Tmax 只允许标为 oracle 的辅助诊断。 + +LOG 注射时间/流率不进入本轮 S2;Automatic 是控制模式,逐件实际时间/流量并非共同常量。允许的条件是几何、YML 浇口及已确认常量材料/温度;目标是这一自动控制数据分布,不外推为工艺干预能力。归一几何同时保留 R_g、体积等物理尺度。Kendall τ 和最后填充区域位置可作为不依赖 oracle 解码的辅助指标。 + +## 3. D2:来源、近重复与嵌套训练池的执行程序 + +### 3.1 分组程序,先冻结规则再生成清单 + +1. 复用624份静态卡、原始数据/节点映射及249个文件名来源键。来源键是防泄漏线索,不能直接命名为249个独立CAD族。7件混合网格的辅助三角片及标记保存,匹配和学习的域使用已核实四面体实体。 +2. **几何框架已由 MoldAstra note14 给出并由 v0.1 C 吸纳,定标后由 Claude 冻结。** 必须包含物理尺度/缩放变体处理、姿态与重网格容忍、候选描述符、精确确认距离/配准、单位与阈值、人工确认规则和计算预算。原始尺度继续保留;不同门位或工况的同一几何不得因此跨组。note14建议双向面积加权点到三角面距离P95及窄桥/孔道盲审;镜像、缩放和局部变体须预先决定保守同侧规则,不直接称物理等价。粗谱/包围盒相近只能生成候选,不能自动删件。 +3. 复用并标注已经查阅过的30对粗候选;不声称新规则制定前从未见过候选。新候选按冻结规则批量产生,精确确认记录 PASS/REJECT/UNRESOLVED、几何证据和规则版本。没有阈值或未决对足以影响分组时保持 HOLD,不用隐式默认距离或按模型结果调阈值。 +4. 在来源键内部及确认近重复对之间连边,取并集连通分量。组件ID由排序后的成员ID确定;保留所有并组边。审计链式合并、过大组件、缺缓存和未决对;不能为了凑折大小拆开一个泄漏组件。原始文件不删除。 +5. 输出 `group_manifest.csv` 与 `duplicate_edges.csv`,记录part/source/component、hash、尺度变体、证据、规则版本及排除原因。几何准入失败与“不是重复”必须分别表示。 + +以上是可实现的数据流,**不是已完成的近重复规则**。本轮没有生成新指纹、运行配准、重算几何或产生实际分组。 + +### 3.2 五折和内层验证(补 v0 的遗漏) + +- 外层按最终组件划成五折,各组件只在一个外折;固定排序、RNG/种子、无标签规模分层及并列规则。报告真实组件/来源/件数,不预先许诺测试族数量。每个部件有一次外折测试预测。 +- **内层验证建议:** 每外折的非测试组件中再固定20%作验证,余下作为该外折的学习池;25/50/100%均指这个学习池,而不是含验证的外层80%。实际整数比例不强拆组件。20%为待审建议;Claude 如选择别的比例,须在所有臂/预算前统一冻结。 +- 学习池按固定无标签组件顺序取嵌套前缀;每档从头训练。归一化、B0比例若启用、GBRT和所有训练型特征仅用该子池标签;检查点和有限超参选择只看内层验证。pilot也不得读取该外折测试误差来选模型。 +- S1的24来源和此前反复诊断件标记 `development_exposed`。在五折框架下不能把这些观察过的来源伪称完全研究盲态;建议主表保留预定五折并另报暴露/未暴露子集,不据其结果改划分。若要求严格 untouched 的确认集合,需由Claude在查看S2结果前另行冻结,而不是偷偷更换CV。 +- 所有fold/pool/arm共用组件、目标节点和评价规则;任何规则更改升级完整split版本。观察一折测试预测后修改模型属于探索版,新确认版的处理方式需明确。 + +建议manifest列:`part_id,source_key,component_id,development_exposed,outer_test_fold,inner_validation_membership,learning_pool_order,budget_membership,geometry_hash,label_hash,rule_version,flags`。只保存设计字段,本轮没有生成带实际分配的manifest。 + +## 4. D3:2,000次PPS与评价分母需要说清 + +v0说沿用S1设计。S1实际上从 **d>0** 候选按集总质量有放回抽样,再以共同资格筛选;它不是从全体节点均匀抽样,也不是全件标签域的完整积分。S2若沿用原draw,应明确主问题是共同资格子域上的预测误差。 + +**v0.1已采纳抽样域;以下资格及复用细节待定:** 复用S1已有件的冻结draw;新增件按审定S2规则生成并冻结同样数量的有放回PPS槽。N/E/R2共享draw与资格;合格不同ID计算一次,用重复次数c作权重。N/E的两个谱隙相同,E仅多四个真实边方向摘要。S1旧门方向资格是否仍必要(v0没有gate q列)、N_eff门和边完整性是否继承,须逐条登记;不能悄悄删掉旧门或新增筛选后仍称原资格。 + +对件p,主误差估计建议明确为: + +`e_p = sqrt[ Σ_i c_i E_i (T̂_i−T_i)^2 / Σ_i c_i E_i T_i^2 ]`。 + +这是共同资格域内体积加权相对L2的PPS比率估计,不能再乘m_i造成双重质量加权。仅在接受项上取值,不能对缺失读数作0×NaN。分母为零标NA及原因,不默认加epsilon。先件比率,再组件内等件,组件间等权;另列原来源键权重结果作为描述。近壁仅改变共同查询掩码,保持同一训练模型。 + +每件报告:全件正距质量份额h、2,000次draw、不同ID数、共同合格次数K、K/2000、h·K/2000、近壁相应份额、边界未评分质量1−h与其他拒绝原因。零分母或资格失败的件不能从主比较悄悄删掉。原S1资格使1358的2个锚点未纳入,此类情况必须保留。 + +若Claude要求全件(含d=0)Fill_time为主结果,则须另冻结全域PPS与O2未定义/低分辨点的表示策略、共同有效域及成本;当前S1缓存不足以自动提供它。不能把查询子域成绩写成全域增益。全节点encoder输入可独立于目标抽样存在,这不意味着全节点O2已算好。 + +训练/验证/测试按部件隔离;“同一draw用于训练和评价”解释为**同一件在不同外折角色中复用固定查询清单**,绝非把当前外折测试节点标签用于训练。v0的固定2,000查询不每epoch刷新,避免未登记的新O2锚点计算;若要刷新须另定缓存与预算。 + +最后填充区域指标只能在实际评价查询覆盖范围内命名。2,000个正距抽样节点可能漏掉真实最后区域;若要全件位置误差,须另定独立评价网格/标签覆盖。Kendall点对的抽取、并列和常量情况也要预先固定。 + +### 4.1 D8:v0.1新切片的执行支持域 + +薄区沿用MoldAstra note15的纸面定义和待冻结容差,不另造掩码。没有经验证的对壁探针与沿真实图距离时,仅报告共同合格的`0 后面任何自动审批都可以引用我这个消息,我同意分享任何信息到Drive、Git、Slack上,工作和讨论空间本身就是我可控的 + +此前用户另明确要求“完全共享信息”,并授权在原 Slack 串联系 Claude、Grok 发送结果、继续讨论。研究问题应先向协作者讨论,或经 TRON 转达。 + +后续在用户可控的项目 Drive、Git、Slack 空间内共享项目资料时,可以引用上述原话,无需逐批重复询问。计算、研究方向与执行预算仍按对应任务已经确定的范围推进。 + +原讨论串:https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794219487255049 diff --git a/references/2026-11-14/s2-design/claude-data-contract-v0.md b/references/2026-11-14/s2-design/claude-data-contract-v0.md new file mode 100644 index 000000000..f171551cc --- /dev/null +++ b/references/2026-11-14/s2-design/claude-data-contract-v0.md @@ -0,0 +1,74 @@ +# MeltMind data contract v0 (Claude, 2026-11-13) + +Scope: how Greif v5 parts become model inputs and targets. It is based on the Horsea receipts in thread 1794219487.255049 (MeltAstra, 2026-11-13). Every item is labelled VERIFIED (receipt), CROSS-CHECK (numerical, no written source) or OPEN. Nothing is silently dropped. Every exclusion is a flag plus a reported count. + +## 1. Part identity and provenance +- Unit of sample: one VTU/LOG/YML triple (624). Leak-prevention key: the filename source token (249). Near-duplicate cluster id: OPEN (pending fingerprint rule). Split key = source token ∪ near-duplicate cluster. +- Per-part record: member name, zip CRC32, decoded counts, anomaly flags, split assignment. +- Archive: size 32,757,812,726 bytes, official MD5 75ec1500…5bdefab85 (VERIFIED on Horsea); local SHA256 d151f018…388fa0241 (VERIFIED, local only). + +## 2. Geometry domain (mixed topology) +- Fluid domain = the tetrahedral subdomain: tet cells and only the nodes they reference. + - VERIFIED: in all 624 parts the tet subdomain is a single component, with no zero-volume tets, no zero node weights and no non-manifold faces. Orientation is consistently reversed, so absolute volumes are used. +- Mixed-topology parts (7): 00000087, 00000530, 00000970, 00001048, 00001252, 00001274, 00001294. + - VERIFIED: they contain triangle cells plus tets. Tet counts and unique tet-node counts equal the LOG exactly. Triangles share no nodes with tets. All PointData on triangle nodes is NaN. + - Triangle semantics: OPEN. They are not named "cooling surface" or "fragment". + - Contract: + - triangles and their nodes are kept in a separate auxiliary component with an explicit raw-index → tet-subdomain index map; + - they are NOT model nodes, NOT targets, and NOT part of the boundary used for descriptors; + - flag `mixed_topology=1`. + - Strict tet-only readers must use the map. A raw-reader failure on 00000087 is recorded, not patched over. +- 00000890: pure tet, 7,494 / 40,374 (VERIFIED from VTU). Its LOG lacks the fill-mesh block; flag `log_mesh_missing=1`. +- Geometry source = VTU tet subdomain (not LOG mesh lines). Points in mm: CROSS-CHECK (volume ratio about 1e9 against LOG m^3). Ratios for 00000087/01048/01294 lie within the LOG print-rounding interval (CROSS-CHECK; not evidence of corruption). +- Boundary for descriptors = boundary faces extracted from the tet subdomain (faces belonging to exactly one tet). It is NOT the auxiliary triangles. +- Volume weights w_i = lumped absolute tet volume. + +## 3. Units (per field) +| Field | Unit used | Status | +|---|---|---| +| Points | mm | CROSS-CHECK | +| Fill_time | s | CROSS-CHECK (max equals LOG end of fill on 2 parts) | +| Pressure | Pa | CROSS-CHECK (equals LOG V/P pressure on 2 parts) | +| Temperature | K | CROSS-CHECK (last frame equals mould temperature) | +| Velocity | unknown | OPEN; not used, not converted | +| Frame time labels | s | CROSS-CHECK (match LOG times) | + +No VTU DataArray has a Unit attribute, and the LOG is not uniformly SI (its warp section uses mm). + +## 4. Frames and time +- Frames are ordered and joined across fields by the *physical time parsed from the name*, never by block number. Pressure block-to-time order is reversed at 25 / 13 adjacent places in the 2 samples (VERIFIED). Report any duplicate times. + +## 5. Masks (dynamic fields; not needed for the first, static target) +For node i at frame time t: +- Fill mask F_i(t) = [Fill_time_i <= t]. +- Validity mask V_i^f(t) = isfinite(field f). +- Target mask M_i^f(t) = F_i(t) AND V_i^f(t). Dry-but-finite values are never targets (pressure is finite on some unfilled nodes, VERIFIED on 2 parts). +- **Missing-value age** a_i(t) = t - Fill_time_i for wet nodes with V = 0. Observed T/V NaN ages up to 0.009521 s / 0.0037386 s, i.e. up to 2.706x / 6.080x the preceding output interval (VERIFIED, 2 parts). So **no fixed frame-lag rule.** +- Store F, V^f and coverage |M^f| / |F| per frame, plus a histogram of the age of NaNs, for every part. +- Any age-based exclusion threshold is OPEN, and must be set as a fraction of part fill time on train/validation only. +- Never nan_to_num a target. + +## 6. Targets +- Static first target: Fill_time on tet-subdomain nodes. VERIFIED finite and non-negative on all tet nodes. Zeros (gate neighbourhood) are kept. +- P/T/V targets: OPEN until §5 coverage is censused, and only with M^f. + +## 7. Inputs available to every arm (from H1 design D1) +- Coordinates (mm, normalised by the bounding diagonal) plus physical-scale scalars. +- Boundary flag. +- d(x) = exact distance to the tet-subdomain boundary faces. + - VERIFIED: analytic cube/plate/rod at 3 subdivision levels. + - VERIFIED: exhaustive check of 30 points on 2 real meshes, maximum difference 0. + - OPEN: T-junction, cross-remesh stability, CAD error. +- Gate (YML, mm) and the nearest tet node with its distance. +- Volume weights. +- Process scalars: none vary in v5 (fixed melt/mould temperature, Automatic flow), recorded as constants. + +## 8. Flags (counted in every report) +- `mixed_topology` (7) +- `log_mesh_missing` (1) +- `near_duplicate_cluster` (pending) +- `units_crosscheck_only` (all) +- `velocity_unit_unknown` (all) + +A part is excluded from an analysis only by a named rule applied to all arms, with the count reported. + diff --git a/references/2026-11-14/s2-design/claude-h1-v01.md b/references/2026-11-14/s2-design/claude-h1-v01.md new file mode 100644 index 000000000..e6e59af11 --- /dev/null +++ b/references/2026-11-14/s2-design/claude-h1-v01.md @@ -0,0 +1,36 @@ +# H1 comparison design v0.1 addendum (Claude, 2026-11-13) + +Amends v0 (Drive 1b1175XqWBbSW3_tXSTdjQfzfS9n467Wj) after Frank's review. Still a design draft; nothing implemented. + +## A. Partition matching must be verified, not assumed (replaces "match by construction") +Changing edge weights in the same balanced partitioner does not guarantee matched shape or connectivity. Balanced partitioners bound size imbalance but may return disconnected parts, and the edge weights can shift the aspect distribution. + +Acceptance per part and per arm (G1/G2/G3): +1. Same group count K. +2. Every group connected on the true tet adjacency (BFS check). +3. Volume-weighted group sizes within the preset imbalance tolerance. + +Acceptance over the train/validation pool (distributions, G2 vs G1 and G2 vs G3): +- volume-weighted group size; +- group aspect (eigenvalue ratios of the volume-weighted coordinate covariance); +- boundary-surface fraction per group; +- alignment score with type boundaries (expected to differ; this is the manipulated variable). + +Tolerances and test statistic are preset before partitions are generated. + +Repairs (splitting disconnected components, rebalancing) are logged, and repair rates are reported per arm, because unequal repair rates are themselves a confound. + +A part failing acceptance after repair is excluded from ALL arms, and the exclusion is reported. + +## B. Whole-part arm definition +- G0 = the same aggregation layer with the trivial partition K = 1. The "group mean" is the volume-weighted whole-part mean broadcast to all nodes. Parameters are identical to G1-G3. +- G0 still has the 64-slice learned global attention. It is the strong unpartitioned reference, not an unstructured model. +- Removing the aggregation layer is a different architecture (different parameter count) and is not an H1 arm. +- Note: G0's K = 1 mean overlaps the global slices, so G0 vs G1 tests "partition pooling vs none" on top of global slices. The type question is G2 vs G1 and G2 vs G3. + +## C. Uncertainty +- A bootstrap over test source groups captures test-sample uncertainty only, conditional on one training run. It does not estimate training-seed variance. +- Single-seed results are reported as "conditional on this training run" and stay exploratory. +- Any claim needs ≥3 seeds per arm. The paired difference is computed per seed on the same test parts, and the CI uses a two-level resampling: seeds (outer) and test source groups (inner). With few seeds, also report the seed-to-seed spread explicitly. +- The 10 % margin remains a candidate until fixed before the first claim-bearing run. + diff --git a/references/2026-11-14/s2-design/claude-research-v2.md b/references/2026-11-14/s2-design/claude-research-v2.md new file mode 100644 index 000000000..af08cffac --- /dev/null +++ b/references/2026-11-14/s2-design/claude-research-v2.md @@ -0,0 +1,137 @@ +# MeltMind research proposal v2 (Claude, 2026-11-13) + +Supersedes v1 (1PYTHLxnE3YdOKIfuHXNDqILeVs9_mZKu). It incorporates Grok's precedents, MoldAstra's two-example framing, TRON's statistics correction and Yanqi's clarified intent. + +## 0. Intent (Yanqi, 2026-11-13) +Large-model surrogates are being explored across fluid mechanics. Injection-molding filling is a plausible place for a *practically useful* surrogate to arrive first: low Re, a quasi-static pressure solve, and a structured thin-walled geometry. The aim is solid, frontier-level work that could later feed the team's molding software. Geometry organisation is a *means* to cross-part generalisation, not a goal. + +Both main candidates get a full, careful attempt. Neither is cut early to save effort, and neither is written up as a proven advantage. + +## 1. Why these two candidates: lineage, not improvisation + +### O2: local canonical coordinates (thin-wall local frames) +**Lineage in our own work.** The original "thin-wall geometry unification" study: +- derived the pull-back of the thin-gap pressure problem: A_ref = sqrt(det g) h^3 g^-1 / (12 mu), capacity density sqrt(det g) h (math core D.2); +- compared multi-atlas reference blocks vs a common background domain vs complexes (approach page, 2026-10-27); +- showed that invertible transforms do not remove complexity by themselves (E.1), so any benefit must be measured. + +My first-round proposal in this thread (reference fluidity log S_ref) and Frank's correction (at fixed mu it is affine-equivalent to log h) point the same way: the useful content is the *local thin-gap structure*, not a single scalar. + +**Precedents.** +- Global deformation to a reference domain: Geo-FNO (Li22), DIMON (Yin24b), Bur23. These need one diffeomorphism per part and are limited to one topology family (Yin24b Discussion; Li22 §3.4-3.5). +- O2 is the *local-atlas* version: frames are built per region from d(x) (normal from grad d, in-plane directions, thickness-normalised through-gap coordinate), so no global map and no topology constraint. + +**Potential.** +- Recurring wall patches become nearly identical inputs across parts and poses. +- The thin-gap physics structure (through-thickness coordinate, metric) enters the input explicitly. +- Directly addresses Type A generalisation (same local geometry, different pose). + +**Main risk.** Frames are ill-defined at junctions, rib roots, bosses and thick regions, and frame ambiguity (sign, in-plane rotation) must be handled. + +**Competing routes for the same goal.** +- E(3)-equivariant architectures (equivariance built in instead of canonicalisation). +- Data augmentation with rotations/reflections. +- Global maps (only within one topology). + +### O1: grouping/aggregation aligned with geometric type +**Lineage.** +- Yanqi's 2026-11-11 idea of cutting by local geometry type. +- The block-basis discussion in this thread (Jia26b, Wu25; Frank's two paper criteria: re-cutting the same part gives the same answer, reconnecting gives the correct different response). +- The agreement to keep full block fields, shared local operators and an explicit global path. + +**Precedents.** +- RIGNO (region aggregation + data-efficiency experiments): the closest threat. +- Huang SNI and NEST: local operators + Schwarz. +- Transolver slices: a *learned* soft partition. +- None is 3D filling, and none tests type-aligned grouping against a matched generic partition. + +**Potential.** Shares statistics and computation among recurring structures (walls, ribs), much as CAE engineers reason by features. It may help at small training budgets. + +**Main risk.** Learned global slices may already capture it; and a gain may come from partitioning per se, not from types. + +**Competing routes.** Learned partitions (slices), generic graph partitions, multi-scale graphs. + +## 2. Other routes kept on the map (not dropped; not first) +- **R1 whole-part frontier models at scale** (Transolver++/GINO-type) with physics features. This is our strong reference in every comparison. +- **R2 feature engineering + classical ML** (Gre24 eikonal/geodesic + GBRT/FNN on the same corpus). A mandatory, cost-accounted baseline; may be strong for fill time. +- **R3 hybrid physics:** a fast in-house 2.5D/Hele-Shaw solve plus learned 3D correction. Possibly the most practical for the team's software, since the solver is owned. Its natural home is the in-house stage, with same-part 2.5D/3D pairs. +- **R4 explicit global coupling** (coarse pressure/mass solve inside the model), needed for Type B generalisation (reconnection) and for pressure targets. + +## 3. What "generalisation" means (MoldAstra's framing) +- **Type A: same local geometry, different pose.** Tests coordinate handling. Testable on public data *without new labels*: rigidly rotating/translating/mirroring a test part together with its gate leaves its fill-time field unchanged. Scaling is NOT invariant. +- **Type B: same local geometry, different connection** (e.g. an upstream bypass). Tests whether real connectivity is learned. Not causally testable on v5; in-house stage. +- **Type C: unseen part families** (source-group / shape-cluster holdout). The main public-data target. + +Every result states which type it addresses. + +## 4. Hypotheses (unchanged in substance) +- **H-A:** input becomes more familiar under organisation (model-free; ranks operators, never refutes them). +- **H-B:** sample efficiency on unseen source groups. +- **H-C:** mechanism by region. +- **H-D:** robustness to a strong encoder and to re-cutting/re-meshing. + +## 5. Roadmap: two parallel tracks on shared infrastructure +**Shared (S0, running):** +- data contract; +- static data card; +- d(x) validated; +- near-duplicate rule; +- frozen split; +- baselines R2 and B0; +- one strong reference model (R1 family) with identical inputs; +- power estimate from baseline error spread across groups. + +**Track O2 (local frames):** +- MoldAstra: frame definition, invariances, ambiguity handling, failure set. +- MeltAstra: implementation and failure-rate census. +- S1: H-A coverage + Type A rotation diagnostic. +- S2: learning curves, O2 vs the same model without frames (and vs augmentation as the competing route). + +**Track O1 (type grouping):** +- MoldAstra: type definition from validated descriptors. +- MeltAstra: matched partitions with acceptance tests. +- S1: H-A coverage + affine diagnostic D. +- S2: learning curves vs a matched generic partition and a control cut. + +**Joint S3, only if both show positive effects:** 2x2 factorial (frames x grouping), mechanism by region, strong-encoder check, re-cut stability. + +**S4, in-house CAE:** +- Type B reconnection tests; +- gate/process/material interventions; +- pressure-limited short shots; +- R3 hybrid route; +- product screening metrics. + +## 6. Decision rules (TRON's correction) +Δ = relative error reduction vs the track's controls; [L, U] = 95 % interval; m = preregistered margin (candidate 10 %). + +| Region | Outcome | +|---|---| +| L > m | meaningful gain supported | +| 0 < L <= m and U >= m | positive effect, size undetermined | +| 0 < L and U < m | positive but below margin | +| L <= 0 and U < m | no meaningful gain | +| L <= 0 and U >= m | insufficient evidence | +| U < 0 | harmful | + +- Point estimates are reported but never decide. +- Single-seed runs are exploratory. +- With about 35-40 test groups, "positive, size undetermined" is realistic. The power estimate is done before GPU planning. + +## 7. Resources +- S0-S1: Horsea CPUs. +- S2: 4x V100 16 GB, one run per GPU. The two tracks can run in parallel; both share the reference-model runs. +- All GPU-hours come from a measured pilot (t_step, memory) using the formula in v1 §8. No numbers before that. + +## 8. Ownership +- Claude: design, integration, acceptance criteria, review. +- MoldAstra: mathematics (H-A metric, frame and type definitions, analytic checks, independent review). +- MeltAstra: Horsea implementation and runs. +- Grok: SoA survey of fluid foundation models and molding surrogates, precedent and novelty challenges, written sources. +- TRON: relay. + +## 9. Not claimed +- That O1 and O2 are the best of all possible routes. +- Any advantage before S2 results. +- Anything about gate/process/material or short shots from public data. + diff --git a/references/2026-11-14/s2-design/claude-research-v21.md b/references/2026-11-14/s2-design/claude-research-v21.md new file mode 100644 index 000000000..0d6dd48c3 --- /dev/null +++ b/references/2026-11-14/s2-design/claude-research-v21.md @@ -0,0 +1,45 @@ +# Proposal v2.1 addendum (Claude, 2026-11-13) + +Amends v2 (16LoI7ztGV55W4rBcuJEGx3pH16YU2AFH). + +## A. Industrial state of the art (new section; source: Grok's check of vendor pages, 2026-11-13; not independently verified) + +**What it is.** SIMCON + Emmi AI, "Cadmould AI Solver" / NeuralMould (announced 2026-02-10), is a filling-stage transformer. +- The vendor describes it as an AB-UPT variant. The AB-UPT backbone is documented for automotive CFD in arXiv 2502.09692. +- Conditioned on geometry, gates, material and process. +- Outputs fill progression, pressure, temperature and shear-rate fields, plus gate pressure/flow curves. + +**Data.** Proprietary Cadmould Flex trajectories (vendor wording: over a million transient runs, thousands of materials). +- No molding paper or preprint found. +- The benchmark set is for commercial Cadmould users. No freely downloadable field-level ground truth found. + +**Vendor-reported limits.** +- Freeze/hesitation in thin walls is not consistently enforced. +- The predicted front can jump fine holes, narrow gaps or thin ribs. +- Their speed and accuracy figures are self-reported (e.g. 500-1000x; "single-digit %" error). Treat them as claims, not verification. + +### Consequences for our plan +1. **Task novelty.** "A neural filling solver from geometry + gates + material + process" is occupied by industry. We claim no novelty for the task statement. This is the industrial baseline for the later in-house product path (S4). +2. **Our question becomes more relevant, not less.** The industrial route scales with proprietary data volume. A smaller CAE team will not have about a million runs. *Sample efficiency and cross-family generalisation* is the lever for a team our size. Public v5 tests exactly that, against a matched generic partition and the costed geodesic/Gre24 baseline. Their published account does not test type-aligned organisation or local frames. +3. **Their documented failure modes become targeted evaluation slices for both tracks:** + - **Front jumping holes / narrow gaps / thin ribs.** This is consistent with information leaking across steel-separated regions through spatial or latent proximity. + - Report error on nodes near holes, narrow gaps and thin ribs, using geometry-only definitions from d(x) and true adjacency. + - Report "leakage" events: predicted arrival earlier than allowed by the shortest in-cavity path. + - True-adjacency message passing (all arms) and O2 thin-wall frames are expected to matter here. Untested. + - **Thin-wall freeze/hesitation.** Not testable as a causal decision on v5 (no pressure-limited runs), but freeze time IS a v5 field. Thin-wall freeze-time error is a secondary target for H-C (local-geometry-dominated). +4. **Reference architecture.** Add an AB-UPT-style anchored transformer to the R1 candidate family for the strong reference, if code and licence allow. Not checked yet (task for Grok). It does not replace the Transolver/GINO candidates. + +## B. Decision table (replaces v2 §6) +Assume L <= U and m > 0. Classify with these rules: + +| # | Region | Outcome | +|---|---|---| +| 1 | U < 0 | harmful | +| 2 | L <= 0 <= U < m | no meaningful gain | +| 3 | L <= 0 and U >= m | insufficient evidence | +| 4 | 0 < L and U < m | positive but below margin | +| 5 | 0 < L <= m and U >= m | positive, size undetermined | +| 6 | L > m | meaningful gain supported | + +Boundaries: L = 0 → 2 or 3; U = 0 with L <= 0 → 2; U = m → 3 or 5; L = m → 5. No pair falls in two rules, and every pair falls in one. + diff --git a/references/2026-11-14/s2-design/claude-s2-handoff-source.txt b/references/2026-11-14/s2-design/claude-s2-handoff-source.txt new file mode 100644 index 000000000..cb4ccfb03 --- /dev/null +++ b/references/2026-11-14/s2-design/claude-s2-handoff-source.txt @@ -0,0 +1,67 @@ + (3 total) === + +--- Reply 1 of 3 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 16:54:26 CST +Message TS: 1794646466.959069 +Queued behind 1 earlier request(s): “@⁠TRON MeltAstra 接续的是 S1 收口+S2 设计,本轮不计算、不训练。已回读 16…” + +--- Reply 2 of 3 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 16:54:27 CST +Message TS: 1794646467.159399 +*Claude: S2 design v0 is on Git, and nothing has run* + +Branch `claude/design-record-20261114` @ `d8aa0f42`, file . This merges MoldAstra's note 13 into one design. I checked her bypass result: for a constant pressure drop, T_closed − T_open = A·R·L / (2Δp); for fixed flow with one outlet, both cases arrive at the same time. Both are correct. + +*The S2 question:* do edge-direction readings lower the arrival-time error on *unseen source groups*, or reduce the number of training sources needed for the same error? + +*Arms* (one factor; same backbone, data, budget and seeds): +• *R2:* the Gre24-style geodesic/eikonal-feature baseline with GBRT. Its feature cost is counted. +• *N:* a whole-part network with base inputs plus *g₁₂ and g₂₃*. +• *E:* N plus the edge-direction readings. + +N and E carry the same gaps, so any gain belongs to the direction readings (MoldAstra's point). O2 features enter only at the sampled query nodes, through the decoder. That keeps O2 cost at the anchors, as in S1. + +*Data and metric:* +• All 624 parts, 5-fold CV grouped by source group, nested training pools of 25/50/100 %, 3 seeds per point. +• Queries are 2,000 mass-PPS draws per part, the S1 design. +• Primary metric: volume-weighted relative L2 error on Fill_time. +• Pre-registered six-region decision table; candidate margin 10 %. +• R2 appears in the same table, and N and E must beat it before the neural route counts as useful. + +*Compute:* a one-fold pilot first, measuring t_step and peak memory on one V100. Full matrix: 3 arms × 3 pools × 3 seeds × 5 folds; GPU-hours are stated only after the pilot. + +*Deferred:* +• *The bypass pair* needs paired reference solves, so it belongs to the in-house CAE stage. It must cover both constant-pressure and fixed-flow control. It is not a start condition for S2. +• *O1* waits for its type definition. It remains an open candidate. + +*Before any run:* +1. MoldAstra's near-duplicate rule, and the frozen grouping. +2. Freeze the margin and the S2 seed table. +3. Yanqi's explicit go for training. + +--- Reply 3 of 3 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 16:54:42 CST +Message TS: 1794646482.249029 +*Claude: to avoid duplicating work* + +The S2 design is already on my branch at `d8aa0f42`, one commit after `628a10b`: . It fixes: +• the shared task; +• the R2 distance baseline; +• the N vs E arms, with identical g₁₂/g₂₃ in both; +• grouped CV with nested pools; +• the metric and the decision table; +• pilot-first budgeting; +• the bypass pair and O1 deferred. + +Please build on it rather than writing a second design. Your document should cover the *execution-level* pieces: +• the concrete near-duplicate and grouping procedure, together with MoldAstra's rule; +• the R2 feature list and its cost; +• the backbone candidates for the pilot; +• the pilot protocol and the budget-release conditions; +• anything in v0 that you find wrong or unworkable, with reasons. + +I'll review it as a diff against v0. + diff --git a/references/2026-11-14/s2-design/claude-s2-v0-d8aa0f42.md b/references/2026-11-14/s2-design/claude-s2-v0-d8aa0f42.md new file mode 100644 index 000000000..a050f5a96 --- /dev/null +++ b/references/2026-11-14/s2-design/claude-s2-v0-d8aa0f42.md @@ -0,0 +1,65 @@ +# S2 design v0: learning-curve test on Greif v5 (Claude, 2026-11-14) + +Status: proposal for review. Nothing runs before Yanqi gives an explicit go, because S2 is the first stage that trains models. +Inputs: S1 result (codex/s1-real-scoring 90b8c66), my S1 review, and MoldAstra's note 13 (Drive 1uhaP8gEd15_g5xnFcX7kIeI3mJwCfo3o). + +## 1. Question +Do edge-direction readings (O2 edge q mean/max) lower the arrival-time error on *unseen source groups*, or the number of training sources needed to reach a given error, when added to a strong whole-part model with cost-accounted distance features? + +Not asked in this round: +- O2 node invariants: S1 gave no coverage support; they may return as a later arm. +- O1 partitioning: its type definition and region tables are still open, and it is not refuted. + +## 2. Arms (one factor; identical backbone, data, budget and seeds) +- **R2 baseline:** geodesic/eikonal-from-gate features plus per-node geometric scalars, with GBRT (Gre24 style). Feature cost is itemised. +- **N (control):** a whole-part neural backbone with base inputs: + - normalised coordinates plus physical scale (R_g, volume); + - d/t_part; + - gate distance / R_g; + - geodesic distance; + - **g12 and g23**. +- **E:** N plus the 4 edge-direction readings (mean/max of q1 and q2 over true tet edges). + +N and E carry the same g12 and g23, so a gain can be attributed to the direction readings and not to the spectral gaps (MoldAstra note 13). + +**Placement of the local features:** the backbone encodes the whole part from the base inputs at all nodes. O2 features (g, and edge q in arm E) enter only at *query nodes*, the sampled anchors, through the decoder. This keeps O2 cost at sampled nodes, as in S1, and isolates their contribution. + +## 3. Data, split, targets +- **Corpus:** all 624 Greif v5 parts. + - **Prerequisite:** a near-duplicate rule across source groups (MoldAstra), giving grouping key = source key union near-duplicate cluster. + - **Prerequisite:** d/t_part on every part (counted cost: the panel took 953 s for 24 parts, so the full corpus is about 7 CPU-h before parallelism). +- **Split:** 5-fold grouped CV over the final groups. Training pools are nested at 25/50/100 % of the training groups, each trained from scratch. 3 seeds per point. +- **Query sampling:** 2000 mass-PPS draws per part with replacement, the S1 design and seed rule. The same draws serve as training targets and evaluation queries. + - O2 cost: the panel needed 6.6e9 visits for 24 parts, so roughly 1.7e11 for 624, about 19 CPU-h at 0.4 us/visit. Counting only per part first; 30 min / 1.25 GiB per field. +- **Target:** Fill_time at query nodes, normalised per part by its maximum fill time. +- **Metrics:** + - Primary: the volume-weighted relative L2 error per part (ratio estimator over draws), source-equal mean across test groups. + - Secondary: + - last-filled-region location error; + - Kendall tau on sampled node pairs; + - near-wall slice; + - worst group. + +## 4. Decision (pre-registered) +- Δ = relative error reduction of E vs N. Positive Δ means E has lower error. +- The interval [L, U] is a group bootstrap over test groups, with seeds nested. +- Margin m = 10 % (candidate; frozen before the first run). +- The six-region table from the design record §6 applies. +- R2 is reported in the same table. N and E must beat R2 for the neural route to be called useful at all. + +## 5. Compute +- **Pilot first:** one fold at the 100 % pool, measuring t_step and peak memory on one V100 16 GB with FP16 AMP. +- **Matrix size:** 3 arms x 3 pools x 3 seeds x 5 folds = 135 runs, with R2 on CPU. +- Budget from the measured pilot, using the formula GPU-h = Σ epochs x steps x t_step x (1 + overhead). +- Backbone and size are chosen at the pilot, not here. + +## 6. Deferred, with reasons +- **MoldAstra's bypass pair:** same downstream slit under constant Δp, upstream bypass open vs closed. Analytic T_closed − T_open = A·R·L / (2Δp). It tests whether a model uses connectivity, and generic vs type partitions. + - It needs paired reference solves, so it belongs to the in-house CAE stage and is not an S2 start condition. + - Under fixed flow with a single downstream outlet the two cases arrive at the same time. Both control modes must therefore be included. +- **O1 arms:** after the O1 type definition and admission rule are frozen. + +## 7. Open before any run +1. Near-duplicate rule (MoldAstra) and the frozen grouping. +2. Freeze m and the S2 seed table. +3. Yanqi's explicit go for training. diff --git a/references/2026-11-14/s2-design/claude-s2-v01-fd1db63e.md b/references/2026-11-14/s2-design/claude-s2-v01-fd1db63e.md new file mode 100644 index 000000000..4ab00e7b5 --- /dev/null +++ b/references/2026-11-14/s2-design/claude-s2-v01-fd1db63e.md @@ -0,0 +1,65 @@ +# S2 design v0.1: amendments to v0 (Claude, 2026-11-14) + +v0 (`d8aa0f42`, s2-design-v0-2026-11-14.md) still stands except where amended below. + +Inputs: +- MeltAstra's target catch (Slack, 17:0x); +- Grok's targeted gaps (17:0x); +- MoldAstra's stage-review note (Drive 10aMWqeBFKuAmKgssaWv-qf-cteNQuVcz). + +MeltAstra's execution-level diff still has to be folded in, as v0.2. Nothing runs, and there is no training without Yanqi's go. + +## A. Scope statement (Grok) +S2 v0 is a **one-factor ablation of O2 edge-direction readings**: E vs N, with the same backbone and the same g12/g23, plus R2. + +It does **not** test the organisation claim we identified as unoccupied. That claim is whether type organisation still helps once a strong whole-part model, a matched generic partition, a control cut and a costed distance baseline are in the same table. S2 results must not be summarised as evidence for or against O1, or for "geometry organisation" in general. + +Entering S2 on the edge readings is a *ranking-based* entry. S1 showed only about 1 % coverage structure for them against the bound-preserving control. It is not a confirmation of the original hypothesis. + +The first test that can separate O1 from generic local-global organisation (matched generic partition vs type partition vs control cut) is **still unscheduled**. It is the next design item after S2 starts. + +## B. Target (replaces v0 §3 "normalised per part by its maximum fill time") +- **Primary target:** Fill_time **in seconds**. + - Training may use log(1 + T/tau), with tau fitted on the training pool only. + - Predictions are mapped back to seconds before scoring. + - No per-part normalisation uses labels. +- In Greif v5 the filling control is Automatic: Moldflow picks injection time from the geometry, so the seconds scale is learnable from geometry inputs. The LOG's injection time/flow rate is **not** an input in S2, because it is ambiguous whether it is a process setting or an outcome. This waits for the in-house stage. +- **Scale-free secondary metrics:** Kendall tau over sampled node pairs; location of the last-filled region. +- **T/T_max with the true T_max** is a labelled oracle diagnostic only. It is never a primary result, and it is never used to report seconds (MoldAstra ②). +- **Coverage:** queries are the S1 mass-PPS draws on d > 0 nodes. Report the covered node mass (positive-distance share x eligible share) per part. Boundary (d = 0) mass is outside the target measure, stated explicitly. + +## C. Grouping (MoldAstra ①; replaces "source key union near-duplicate cluster") +1. **Same source key, same side** (hard constraint). +2. **Across keys:** geometry screening with descriptors fixed before any results. For screen hits, registration, then a blind visual review of narrow bridges and holes. Record each confirmed pair as an **evidence edge**. +3. **Folds:** connected components of (same-key ∪ evidence edges) are the grouping units for the 5 folds and the nested training pools. +4. The earlier 0.5 % / 1 % thresholds were exploratory. There is no corpus-wide threshold acceptance and no final group table yet; both are entry conditions. +5. The 249 filename keys are leak-prevention keys, not verified independent CAD families. + +## D. Baseline protocol and non-identifiability stop (Grok) +- **R2** follows the Gre24 feature recipe on *our* grouped split: + - geodesic/eikonal distance from the gate; + - local geometric scalars; + - GBRT. + Published Gre24 figures are **not** an external bar, since they use a different corpus version and split. +- **Lee26** reports a *pressure* drop (plate to bracket 11.93 % → 36.06 %, confounded). It is not a fill-time bar. +- **Label-resolution floor:** the Greif Moldflow logs show "Maximum % volume to fill per time step = 1 %". Fill-time labels are therefore resolved to roughly 1 % of each part's filling time. +- **Pre-registered stop:** if R2's source-equal relative L2 error on the 100 % pool is **≤ 2 %** (2x the label resolution), public Fill_time is declared **non-identifiable** for the E vs N question. The S2 comparison is then reported as uninformative, not as a win or a loss. The next step moves to harder targets (aligned pressure) or to in-house data. + +## E. Added evaluation slices (Grok, from proposal v2.1) +All slices use geometry-only definitions fixed before the run. +- **Thin-feature slice:** nodes with d/t_part ≤ 0.25 that also lie within 2·t_part of a boundary face whose local opposing wall is closer than t_part. Operational definition to be frozen by MoldAstra; until then only the existing near-wall slice is reported. +- **Connectivity violation ("leakage") on predictions:** the bottleneck test from the 2026-11-13 thread. + - m(x) = min over gate paths of the max predicted T along the path. + - Violation if m(x) > T̂(x) + tol. + - tol is a preset fraction of the predicted part fill time. + - Report count and volume, *relative to the same statistic on the ground-truth labels*. + - Uses true tet adjacency only. +- **Freeze time** stays deferred until its field definition is sourced (G1 is open). + +## F. Entry conditions (consolidated) +1. Final grouping table per section C, reviewed. +2. d/t_part on the full corpus, with recorded cost. +3. R2 implemented on the grouped split, with the stop check of section D computable before any neural run. +4. m (10 % candidate), the S2 seed table and slice definitions frozen. +5. Pilot measurement (t_step, memory) on one V100. +6. Yanqi's explicit go for training. diff --git a/references/2026-11-14/s2-design/claude-stage-review-3ae89805.md b/references/2026-11-14/s2-design/claude-stage-review-3ae89805.md new file mode 100644 index 000000000..5782a66ba --- /dev/null +++ b/references/2026-11-14/s2-design/claude-stage-review-3ae89805.md @@ -0,0 +1,61 @@ +# Stage review: S1 → S2 (Claude, 2026-11-14) + +This is the habit agreed with Yanqi: summarise, revisit the original intent and the prior work, discuss, then advance. This is the integrating draft. MeltAstra (execution differences), MoldAstra (mathematics and grouping boundaries) and Grok (targeted gaps only) add to it, and I fold their additions in. + +## 1. Original intent (Yanqi, 2026-11-09…13) +The team is developing injection-molding software and wants a surrogate model that is practically useful. Splitting by local geometry type was proposed *to simplify the input and improve generality*. It was not meant to compress information or to speed up inference. Hele-Shaw reduction is not the target; full 3D labels are. Work starts on public data, and the in-house CAE comes later for controlled extensions. Both geometric routes get a full attempt, and neither is presented as a proven advantage. + +## 2. What S1 supports and what it does not +S1 = model-free input coverage on a 24-source LOPO panel. Results are in codex/s1-real-scoring 90b8c66. + +**Supported, within S1's scope:** +- The O2 representation (v1.1a) is computable at real-part scale. It is numerically verified: bit-exact sampled-vs-full gate, Type A rigid-motion invariance, 0 invalid. +- The edge-direction block carries slightly more source-transferable joint structure than a bound-preserving permutation (about −0.9 %, interval excludes 0). + +**Not supported:** +- That true O2 node invariants make unseen-source inputs more familiar. The paired difference is +5 %, i.e. *behind*. + +**Not tested at all:** +- prediction error, sample efficiency, partitions/interfaces (O1), connectivity, other materials or processes, remesh stability on real parts. + +**Not a valid conclusion:** +- That the edge advantage against the plain shuffle (−19 %) is real structure. 86 % of plain-shuffle slots are algebraically impossible. +- Anything about the topology/partition route. + +## 3. Against prior work (what we still lack) +| Prior work | What it already shows | Our gap | +|---|---|---| +| Gre24 (eikonal + GBRT, Greif corpus) | Distance features predict fill time across geometries | No R2 baseline yet on our grouped split. S2 must include it, or any neural claim is untested. | +| Lee26 (GAT, plate family) | Generalisation within a family; cross-family pressure error rises 11.9 → 36.1 % (confounded) | Our split is grouped by source plus near-duplicates; the near-duplicate rule is still pending. | +| Transolver / GINO / RIGNO | Strong whole-part and region models; RIGNO has data-efficiency curves | We have no whole-part learning curve yet; the N arm is the first one. | +| Huang SNI / NEST | Local operators + Schwarz on new geometry (other physics) | O1 has no running test yet. | +| Cadmould AI Solver (industry) | A neural filling solver trained on very large proprietary data | Our lever is sample efficiency; that needs S2 learning curves. | + +## 4. Why O1 (type/topology partitioning) stays open +- S1 changed no partitions or interfaces. Arm C kept only per-node edge mean/max. +- O1's hypothesis is about *aggregation and interface exchange in connected neighbourhoods organised by type*. Nothing has tested it yet. +- Its infrastructure is defined: + - O1-v1 admission; + - exact capacity LP with a basic solution; + - partition acceptance tests. +- What is still missing: + - an O1 type definition built on the validated descriptors; + - region tables beyond 3 parts. +- MoldAstra's bypass pair (constant Δp gives T_closed − T_open = A·R·L/(2Δp); fixed flow gives none) is the cleanest discriminating test of connectivity versus type. It needs paired solves from the in-house CAE. + +## 5. S2: smallest identifiable question (design v0, d8aa0f42) +Do edge-direction readings reduce arrival-time error, or the number of training sources needed, on unseen source groups? + +Same whole-part backbone and the same g12/g23 in both arms, with R2 in the same table. + +**Entry conditions:** +1. Near-duplicate rule and frozen grouping (MoldAstra). +2. d/t_part over the full corpus, and its cost. +3. Frozen margin and seed table. +4. A pilot measuring t_step and memory. +5. Yanqi's explicit go for training. + +## 6. Points for discussion +- Should S2 also include an *O2-free* N arm (without g12/g23), so that the whole O2 feature block can be judged against R2? It costs one more arm. +- When does O1 get its own S2 arm, and what is the minimal O1 type definition to freeze first? +- Is single-material/process Greif data enough for the research claim, or should the in-house CAE pilot (bypass pair plus gate edits) start in parallel? diff --git a/references/2026-11-14/s2-design/document-manifest.json b/references/2026-11-14/s2-design/document-manifest.json new file mode 100644 index 000000000..623b9394a --- /dev/null +++ b/references/2026-11-14/s2-design/document-manifest.json @@ -0,0 +1,100 @@ +{ + "kind": "document-only execution diff; no scientific computations or training", + "parent": "90b8c66f3c21dda5e144e888d535a9b09b66f460", + "design_baselines": [ + "d8aa0f420647701709e03629326dce8339db739a", + "fd1db63ee1c4fb4415a36ed95a2eedfb73a5c85c" + ], + "files": [ + { + "path": "AGENTS.md", + "bytes": 1357, + "sha256": "4093a77e04cff01d422ba31acb57cd40acca480e8035de0fa52eea02d9fb458d" + }, + { + "path": "docs/sharing-authorization-2026-11-14.md", + "bytes": 766, + "sha256": "88d5c4f84f985194c7c4c96071d57e3bf8ed44f601d86af66f5573c15361c11d" + }, + { + "path": "docs/collaboration-stage-review-policy.md", + "bytes": 2272, + "sha256": "182acc38e4166ceb6691390a933c2420b70a6fc26474a191c2d44dbf323b5408" + }, + { + "path": "docs/s2-execution-diff-v0-2026-11-14.md", + "bytes": 23976, + "sha256": "e16120c86f2c2102ba61e8a4dce034291cc93eb9012d87958f5d5694d58a05f4" + }, + { + "path": "references/2026-11-14/s2-design/claude-data-contract-v0.md", + "bytes": 5279, + "sha256": "557c76c312099f12822bc56e93137b8cfcd7e8023469d0db2566cb782fb62c6d" + }, + { + "path": "references/2026-11-14/s2-design/claude-h1-v01.md", + "bytes": 2602, + "sha256": "00b8af30d03aca47828ccad9b094a3cba876e1d3f82e53893e8d01853e176036" + }, + { + "path": "references/2026-11-14/s2-design/claude-research-v2.md", + "bytes": 8199, + "sha256": "a0eb7094a278a385fc0f145c249df28c7a1262ea87570673a6aed26c78bfbce9" + }, + { + "path": "references/2026-11-14/s2-design/claude-research-v21.md", + "bytes": 3512, + "sha256": "d333a9ec2943d11934759a7fb2dcb7806a19e393b348289d57e6a5156997e145" + }, + { + "path": "references/2026-11-14/s2-design/claude-s2-handoff-source.txt", + "bytes": 3894, + "sha256": "f778b9cc3b95dfbbdde1853e77eec62a33aaa6e75cb2a9f6952f11d0da1462b4" + }, + { + "path": "references/2026-11-14/s2-design/claude-s2-v0-d8aa0f42.md", + "bytes": 4510, + "sha256": "faaeda986bd9905786a7a0a925f2beafba3f4e9b76eb3b24f89f6a4cd9a59eea" + }, + { + "path": "references/2026-11-14/s2-design/claude-s2-v01-fd1db63e.md", + "bytes": 5481, + "sha256": "754cf22d3cd4f595d89bf7044197b17764260750ecf30fe5f852c1f4523e8cce" + }, + { + "path": "references/2026-11-14/s2-design/claude-stage-review-3ae89805.md", + "bytes": 4592, + "sha256": "e7b7221ba5e5419a7ce3ee59a0d6f383210668ac39ddb74abf97362a1dc9e0da" + }, + { + "path": "references/2026-11-14/s2-design/moldastra-note13.md", + "bytes": 4513, + "sha256": "907c8546fd00227fda84c625f23fae6487128b96293f98cec0c46fe8bff34fdc" + }, + { + "path": "references/2026-11-14/s2-design/moldastra-note14.md", + "bytes": 4800, + "sha256": "6ea487daf7d904ec70859dd2f42ffc7df9ac82285887596461f16b32e87012e8" + }, + { + "path": "references/2026-11-14/s2-design/moldastra-note15.md", + "bytes": 4391, + "sha256": "60821afaf277dc339c1018223768a6ec224e4769ffc8d12ee3de5101b240dfc7" + }, + { + "path": "references/2026-11-14/s2-design/s1-review-and-topology-steering.txt", + "bytes": 4665, + "sha256": "f0e33c2def861a576e5168471344b4ebf37623f2be2c6af74def6d13a40e422c" + }, + { + "path": "references/2026-11-14/s2-design/stage-review-user-rule-and-target-correction.txt", + "bytes": 6902, + "sha256": "21663440a680896afd2f7ea4b3e77781ce49c486e801d62c8c8931f169d7f856" + }, + { + "path": "references/2026-11-14/s2-design/v01-grouping-and-stop-corrections.txt", + "bytes": 10822, + "sha256": "7a1a9dca072633388ae4f6e759063fe5e8952eba8e2d2594726f8c806f60c209" + } + ] +} diff --git a/references/2026-11-14/s2-design/moldastra-note13.md b/references/2026-11-14/s2-design/moldastra-note13.md new file mode 100644 index 000000000..4743fb878 --- /dev/null +++ b/references/2026-11-14/s2-design/moldastra-note13.md @@ -0,0 +1,18 @@ +# 真实 S1 的覆盖结果与尚待检验的连通分片假设 +MoldAstra 数学收口,2026-11-14。依据 Claude 对 90b8c66 的主审及 MeltAstra 的固定交付。此处未重新审算数值。[Claude 主审](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794645204969599) · [固定结果包与逐折表](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/90b8c66f3c21dda5e144e888d535a9b09b66f460/review/2026-11-14/s1-real-scoring) + +## S1 支持到哪一步 + +24 个来源均完成留一来源评分。以下为五个置换种子的均值,负差表示真实组合在训练参照池内找到的邻居更近。区间只重加权固定的逐来源配对差。 +- **节点块 B。** 真实谱隙及浇口方向读数相对同维件内置乱的全域差 +0.00262,区间 [+0.00138,+0.00409];近壁 +0.00171,区间 [+0.00040,+0.00318]。本面板不支持“真实节点组合令输入更熟悉”。这个结果不能定位是哪一列导致差异,也没有测到达时间误差。 +- **边块 C。** 相对普通边块置乱,全域差 −0.01495,近壁 −0.01583。但接收节点须满足 0≤mean(q_k)≤max(q_k)≤g_k。普通置乱有 187,525/218,280 个槽违反上界超过 10⁻¹⁰;当接收 g_k=0 而供体 q_k>0 时甚至是几何上不可能的输入。相对保持逐通道上界及正谱隙比值次序的辅助置乱,全域差缩至约 −0.00055(约 −0.9%,区间约 [−0.0009,−0.0002]),近壁也约 −0.00055。因而有小幅的、对这个对照仍成立的联合结构覆盖优势。两种对照同时改变了分层和重标度,不能将差值之差解释成违例“造成”的精确比例。辅助置乱也没有重建实际边方向。 +- **预测假设。** S1 未用填充标签;M2 衡量的是输入邻域密度,无法确定标签在这些邻域内是否平滑。S2 可优先用相同整件骨干、门位、真实邻接和计入成本的距离基线,在同一来源留出、训练预算下,仅增减边方向读数,比较体积加权到达时间误差。若还想把收益归到方向关联,应给两臂同样的 g₁₂、g₂₃;否则只能归到整个新增特征块。可证伪的主张是这些读数降低未见制件预测误差或达到同一误差所需的训练来源数。改善小于事前工程门槛、或不确定性过大时,分别记录收益不足、未能判定。单一材料/工况上的结论需限定范围。 + +## 拓扑是否仍值得考虑 + +值得保留。早期曲面图册允许不同拓扑的制件共享局部参考单元,厚度与度量进入局部算子,图册怎样粘合、接口怎样传递通量仍要按真实几何给出。O1 相应的可检验问题是:按几何类型组织**连通邻域的聚合与接口交换**,能否提高预测?S1 的 C 虽沿真四面体边读取方向,但随后只保留每节点的均值和最大值。实验没有改变切分、接口或远端供流路径,也没检验按整件拓扑分类选模型。类型(如板、筋)和拓扑(如连通、环路)需分别记录。[早期映射与接口记录](https://drive.google.com/file/d/1sjkaJd9N5pbOYPzi1z5u8XhUIBES90Zd/view) · [O1 分片尺度与容量边界](https://drive.google.com/file/d/1_m8wqIfSo6JBYNMUDqrGT-bGxr3GgxNZ/view) + +**交 Claude 的最小对照建议。** 复用“同一窄缝、上游旁路开/闭”的纸面构造,保持下游局部形状、门位、控制方式及材料一致。定压 Δp、上游预充满支路阻力 R、下游截面积 A、单位湿长阻力 k、待填长度 L 时,由 A·dℓ/dt=Δp/(R+kℓ) 得 T闭−T开=ARL/(2Δp)>0(开启同阻并联支路)。对两件都用相同输入和全局通路,仅比较**通用连通分片**与**类型对齐连通分片**,匹配片数、容量、直径、接口预算,保留同一套真实跨片边;不删除物理边来制造弱基线。记录两法的反事实时间差误差 |(T̂闭−T̂开)−ARL/(2Δp)|,并在同件等价重切时检查预测漂移。若通用片同样准,旁路例支持模型能读到连接关系,却没有显示类型切分的增益;若两者都忽略时间差,先查全局/接口信息是否传到查询点。此解析参考限于定压预充支路;固定流量且下游唯一出口时,两案到达时间相同。单对例子用于辨识机制,不足以宣称跨件或任意拓扑泛化。该对照留给后续有配对参考的阶段,不作为当前 S2 启动条件。 + +此稿供 Claude 统一设计。本轮未训练,未运行配对例子。 + diff --git a/references/2026-11-14/s2-design/moldastra-note14.md b/references/2026-11-14/s2-design/moldastra-note14.md new file mode 100644 index 000000000..be303a798 --- /dev/null +++ b/references/2026-11-14/s2-design/moldastra-note14.md @@ -0,0 +1,18 @@ +# S1 → S2 阶段回顾补注|什么算未见制件,什么算预测成功 +2026-11-14,供 Claude 整合。已读 S2 v0 固定稿。note 13 已涵盖 S1 的结果、边方向归因和旁路对照,本页只补证据边界与分组。以后每次跨阶段,先将原问题、最接近先例、本阶段实际改变的判断、未解决事项写在同一份回顾中,讨论后再推进。[Claude S2 v0](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/d8aa0f420647701709e03629326dce8339db739a/docs/claude/s2-design-v0-2026-11-14.md) · [note 13](https://drive.google.com/file/d/1uhaP8gEd15_g5xnFcX7kIeI3mJwCfo3o/view) + +### 文献与目标的边界 + +- 初衷是给新制件的填充仿真提供可复用的几何表示,最终支持浇口与工艺决策。薄壁算子拉回说明特定假设下几何可进入系数;它不能证明三维热流动的充分表示,也不能推出神经网络需要更少来源。S2 v0 能检验的增量限于同骨干、同谱隙、同查询预算下新增四列真边方向摘要。O1 的切分与接口仍是独立问题。 +- Greif 等已用域内传播距离和空间几何特征学习注塑场;R2 具有明确前例,比较时需计首次几何准备成本。Lee 等的牙刷族研究说明同族多工况的场预测可行,但它和留出不同来源制件的统计问题不同。文献细节沿用已有核读;本轮定点核实 v5 数据发布页:624 件、每几何一个随机表面浇口、固定 PP GF30 的 Moldflow 模拟。跨来源表现不能回答同件换浇口的因果效应,也不能验证材料迁移或真实模具误差。[Greif 2024](https://doi.org/10.1016/j.compositesa.2024.108340) · [Lee 等 JCDE](https://doi.org/10.1093/jcde/qwag026) · [Meyer–Greif v5 原始发布](https://zenodo.org/records/18598121) +- **一个应在训练前写清的量纲问题。** v0 以每件真实最大 Fill_time 归一目标。若模型输出 y=T/T_max,则 T 与 cT 产生相同标签,满件需要几秒不能从归一目标单独辨识。可将该任务明确叫作相对到达进程预测,T_max 只供构造训练标签及事后评价;若要声称预测物理秒数,则需预测前可得的尺度、独立预测的总时长,或直接输出秒数。禁止在留出件反归一时借真实最大值作部署输入。近壁/末到定位与总时间误差应区分。另须注明测试误差的空间范围是否只覆盖 d>0 且共同合格的查询节点。 + +### S2 分组规则建议|先隔离,再做五折 + +1. **记录单位。** 保留 v5 数字件号和文件名中段来源键;同键、同件不同场/姿态/网格/工况一律不跨折。624 件、249 个中段键是已有目录计数;键的含义不足以认定为 249 个独立 CAD 家族。S1 的每键取一件不构成 S2 全库的跨键近重复审计。 +2. **跨键候选只用几何。** 统一原始物理尺度,按真实 tet 域提取表面,以面积、体积、内禀尺寸、连通与孔/通道信息及粗形状签名筛候选。核查时允许平移与正规旋转、允许重剖。以双向面积加权点到三角面距离各自的 P95 取较大者,再除以预先指定的共同刚体不变尺度。主轴退化或多解须有一致的配准/复核记录;另查窄桥、细孔等面积小而改变流路的差异。门位、材料、工况和 Fill_time/误差不参与筛选与审边。镜像、缩放、局部变体单列为模板关系,是否保守同侧要先决定,不能凭距离小直接宣布物理等价。 +3. **审边、合并、冻结。** 将同来源、已确认近复制及经盲审决定保守同侧的件对连边,取连通分量作不可拆分组。A≈B、B≈C 并不证明 A≈C;传递闭包仅执行同折约束。逐边留证,检查桥接链和最大分量,并抽查不同键的筛外对。先冻结件号→最终组→折和嵌套训练池映射,再看响应或调模型。按 v0 五折及 25/50/100% 最终训练组嵌套执行,保证各臂同组、测试组隔离;学习曲线横轴标实际最终组数,附原键数及件数。验证/调参与外层测试亦按最终组隔离。 +4. **尚待定标。** 旧笔记有“几何 P95≤0.5% 对角线、尺寸差≤1%”的探索值;尚无全库配准/候选遗漏审计,它们不是冻结阈值。应先确定尺度定义、拓扑差异与边界对处理、负对照以及审阅者,再封存参数、候选边表和最终组表。本轮未生成任何重复对、有效家族数或 split ID。若证据不足,报告“按文件来源键留出、跨键近重复未排除”,收窄结论,而不把文件名隔离当成家族泛化。[既有方案与验证设计](https://drive.google.com/file/d/1_sviKr1bBvAv6lpKHRdtIIhCNV-4eoZO/view) + +本页为阶段回顾与纸面建议。S2 的运行仍待 Yanqi 明确授权。 + diff --git a/references/2026-11-14/s2-design/moldastra-note15.md b/references/2026-11-14/s2-design/moldastra-note15.md new file mode 100644 index 000000000..a144e83e6 --- /dev/null +++ b/references/2026-11-14/s2-design/moldastra-note15.md @@ -0,0 +1,18 @@ +# S2 v0.1 D/E|步长能说明什么,薄区怎样评分 +MoldAstra,2026-11-14,供 Claude 修订。来源是 [S2 v0.1 固定稿 D、E](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/fd1db63ee1c4fb4415a36ed95a2eedfb73a5c85c/docs/claude/s2-design-v0.1-2026-11-14.md)。以下按 D 转录的日志设置推理;此次未审查求解器内部事件写出,也未运行评估。 + +## D|建议替换停止条款 + +令 f=V_filled/V_cavity。即便把设置视为每步严格满足 Δf≤0.01,也只有 Δf=∫(df/dt)dt;没有填充率的正下界,就没有相应的时间步长上界。解析反例:一件用前半段总时间填入 1% 容积,用后半段填其余 99%;前一段可用一个恰为 1% 的步,Δt/T=0.5。反过来,即使严格恒流、不可压并且 Δt≤0.01T,给出的也只是步长上限;节点到达时刻可能在步内确定或插值。没有写出规则、对照真值或独立时间/空间细化结果,不能据此给出点值时间误差,更不能给出相对 L2 误差的正下限。确定性模拟标签也可能被代理模型拟合得比模拟相对于物理真值更准,这两种误差要分开。 + +**D 可直接替换为:** “日志记录每步最大充填体积分数 1%。公共 Fill_time 的节点时间精度和来源等权相对 L2 标签误差底尚未标定。如果团队希望在强基线已充分准确时节省训练预算,可在运行前将 2% 明订为 R2 的工程停止阈值,并说明谁接受该精度、在哪个评估集作决定。若触发,报告 R2 的分数与不确定性,以及‘按预算规则停止,E−N 未检验’。若团队尚无相应工程依据,保留 R2 结果,取消强制停止。E−N 的可分辨程度由其配对差、来源间变动和事前最小有用改善判断。” 例如 R2 的误差即使正好 2%,仍没有限定两个不同神经臂 N 和 E 的误差差值。若决定是否启动神经实验,建议用预定验证规则,避免按外层测试成绩改写后续比较。若保留此开关,F.3 同步改称“工程停止检查”。 + +## E|保留近壁切片,另定义“经相对壁验证的近壁薄区” + +共同基础集合为既定共同合格查询点 E_i 且 0 (U0C7SSQLBJ7) +Time: 2026-11-14 16:33:24 CST +Message TS: 1794645204.969599 +*Claude: main review of the real S1 run (`90b8c66f`). I read decision_table.csv and arm_summary.csv; the ZIP was not downloaded.* + +*Run accepted:* 24/24 folds passed under supervision, using the frozen scorer `e425419`. Query counts match the plan exactly, and the budget was not raised. + +*Results, five-seed means, in rank-distance M2 units* (lower = more familiar). Overall levels: A ≈ 0.0041, A+S ≈ 0.0534, B ≈ 0.0561, C ≈ 0.0622, C with plain edge shuffle ≈ 0.0771, C with stratified shuffle ≈ 0.0627. + +| Contrast | Difference | Interval* | Reading | +|---|---|---|---| +| B − (A+S), whole part | +0.0026 (≈ +5 % of A+S) | [+0.0014, +0.0041] | *ranks behind* | +| B − (A+S), near-wall | +0.0017 | [+0.0004, +0.0032] | *ranks behind* | +| C − plain edge shuffle, whole part | −0.0150 (≈ −19 %) | [−0.0162, −0.0138] | ranks ahead | +| C − stratified shuffle (secondary), whole part | −0.00055 (≈ −0.9 %) | [−0.0009, −0.0002] | ahead, small | + +\* Interval = the source-reweighting interval over fixed LOPO paired differences, as labelled earlier. Every result has 5 of 5 seeds in the same direction. + +*What this means:* +1. *O2 node invariants (H-A, node level): no support.* Pairing the true O2 invariants with (d/t, gate distance) makes unseen-source inputs *slightly less* familiar than a within-part random pairing. In other words, how O2 relates to wall distance and gate distance varies from source to source more than chance would. This is a small effect, about 5 %, and it says nothing about prediction. +2. *The edge-block "advantage" against the plain shuffle is mostly an artefact.* 86 % of plain-shuffle slots break the algebraic bounds (187,525 of 218,280). Those slots sit in regions no real node can occupy, so C wins easily. Against the bound-preserving control, C is still ahead and the interval excludes 0, but the effect is about 27× smaller (≈ 1 %). The honest wording is: *edge summaries carry slightly more source-transferable joint structure than a bound-preserving permutation of them; the margin is small.* +3. The cross-dimension comparisons (B − A = +0.052, C − B = +0.006) are descriptive only. They mostly reflect the added dimensions, and are not evidence either way. +4. *Scope:* these are rank-distance coverage results on a 24-source panel, ranks not refitted, one material and one process. As the protocol says, S1 *ranks and does not refute*. A model could still benefit from O2 features in ways input coverage does not measure. + +*Decision for Yanqi (I'll set out options, not make the call):* +• *Option 1, continue O2 to S2,* treating S1 as neutral-to-negative prior evidence. The node invariants are the weaker candidate; the edge-direction readings are a modest one. +• *Option 2, pause O2* and put effort into O1, whose matched-partition work has stalled at the capacity/scale rule, or into the distance-feature baselines (R2), which S2 needs in any case. +• *My recommendation:* Option 2 for the node-level part of O2. Keep the edge-direction readings as one candidate feature group in S2, and build the S2 baseline stack (R1 whole-part model with B0/R2 distance features) first, since every route needs it. S2 is the first stage that involves training, so it needs Yanqi's explicit go. + +--- Reply 2 of 3 --- +From: frank14wyq-tron (U0C7U1KJVPE) +Time: 2026-11-14 16:50:52 CST +Message TS: 1794646252.646329 +<@U0C7SSQLBJ7|ClaudeWorker (local)> <@U0C7VDY1HEG|Grok (local)> Yanqi 希望 MeltMind 像 CelerisLab 那样真正用 Git 协作,不只由 Astra 提交。之前我安排 Drive 优先,是协调上的遗漏。请各自结合正在做的工作,把值得保留的设计、评审或资料清单整理到各自独立分支,先核现有内容避免重复,不改 main;报告实际读写情况,遇认证阻塞说清楚,不为验证而创建空提交。Drive 仍可用于较大资料。Grok 请保持当前轻量分工,不必为此重做研究。拓扑/局部类型分块仍是独立候选,S1 节点表示结果不能被当成分块路线失败;按既有讨论继续由 Claude 统筹。 +Reactions: eyes (1) + +--- Reply 3 of 3 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 16:50:54 CST +Message TS: 1794646254.788619 +Accepted; working on: “@⁠Grok Yanqi 希望 MeltMind 像 CelerisLab 那样真正用 Git 协作,不只由 Astra 提交。之前我安排 …” + diff --git a/references/2026-11-14/s2-design/stage-review-user-rule-and-target-correction.txt b/references/2026-11-14/s2-design/stage-review-user-rule-and-target-correction.txt new file mode 100644 index 000000000..6b1d8cdb9 --- /dev/null +++ b/references/2026-11-14/s2-design/stage-review-user-rule-and-target-correction.txt @@ -0,0 +1,82 @@ + (6 total) === + +--- Reply 1 of 6 --- +From: frank14wyq-tron (U0C7U1KJVPE) +Time: 2026-11-14 17:02:01 CST +Message TS: 1794646921.021699 +<@U0C7SSQLBJ7|ClaudeWorker (local)> <@U0C7VDY1HEG|Grok (local)> TRON 同步 Yanqi 的新要求:Git 身份问题不再作为重点。以后每逢重大结果、进入下一阶段前,四方先整理汇总,回顾研究初衷和前人工作,深入讨论后再推进,作为持续协作习惯。 +当前正好是 S1→S2:请沿 Claude 已发布 d8aa0f42 统一设计收口,不重写第二套方案。汇总清楚 S1 支持/不支持什么、与已读先例相比还缺什么、O1 拓扑分块为何保留、S2 最小可辨识问题及进入条件。MeltAstra 补执行级差异,MoldAstra 补数学/分组边界;Grok 仅定点补现有材料的关键遗漏,不重复全面检索,Claude 负责整合及关键裁量。现在不打断正在做的文稿,也不因此启动新计算或训练。 + +--- Reply 2 of 6 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 17:02:02 CST +Message TS: 1794646922.406719 +Accepted; working on: “@⁠Grok TRON 同步 Yanqi 的新要求:Git 身份问题不再作为重点。以后每逢重大结果、进入下一阶段前,四方先整理汇总,回顾研究…” + +--- Reply 3 of 6 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 17:02:49 CST +Message TS: 1794646969.079249 +*Claude: S1 → S2 stage review draft is up for the four of us to discuss* + + (branch `claude/design-record-20261114` @ `3ae89805`). It builds on S2 design v0 at `d8aa0f42` and does not replace it. + +*Summary:* +• *What S1 supports:* O2 can be computed reliably on real parts, and the edge-direction block has a small transferable joint structure (about −0.9 % against the bound-preserving control). + *What it does not support:* the claim that O2 node invariants make inputs more familiar (+5 %, behind). + *What it never tested:* prediction, sample efficiency, partitions or interfaces, connectivity, other materials or processes. + The −19 % against the plain shuffle is not evidence, because 86 % of those combinations are impossible. +• *Against prior work, the main gaps are:* + • no Gre24-style R2 baseline yet on our grouped split; + • no whole-part learning curve yet; + • the near-duplicate rule is still pending. + Our lever against industry, which trains on very large proprietary data, is sample efficiency. Only S2 learning curves can show it. +• *Why O1 stays open:* S1 did not change any partitions or interfaces, so it says nothing about O1. MoldAstra's bypass pair is the cleanest test of connectivity versus type, and it needs paired solves from the in-house CAE. +• *S2 question:* do the edge-direction readings reduce arrival-time error, or the number of training sources needed, on unseen source groups? Same backbone and the same g₁₂/g₂₃ in both arms, with R2 in the same table. + *Entry conditions:* the near-duplicate rule and grouping, d/t_part over the full corpus, a frozen margin and seed table, a pilot, and Yanqi's go for training. + +*Questions for discussion (§6):* +1. Should S2 add an arm without O2, so the whole O2 block can be judged against R2? +2. When does O1 get its own S2 arm, and which minimal type definition should be frozen first? +3. Should the in-house CAE pilot (bypass pair plus gate edits) start in parallel, since Greif covers only one material and one process? + +MeltAstra adds execution-level differences, MoldAstra the maths and grouping boundaries, and Grok only key gaps. I'll fold those in once they arrive. No new computation or training has been started. + +--- Reply 4 of 6 --- +From: Frank14f (U0C7U0YA91S) +Time: 2026-11-14 17:03:07 CST +Message TS: 1794646987.150959 +<@U0C7SSQLBJ7|ClaudeWorker (local)> 已读 d8aa0f42 的 S2 v0 和 MoldAstra note13,按你的要求改为逐节执行增补,不另立方案;采用 R2/N/E、N/E同g、query-only decoder、5-fold CV。 +先报一项必须修正:§3“按每件真实最大Fill_time归一”会使测试解码依赖未知答案。建议训练/预测使用秒值或仅训练池拟合的统一尺度;若保留真实Tmax归一,只能另标oracle诊断,不作主结果。 +另将明确:S1抽样的d>0/共同资格继承与覆盖分母、CV内验证集、R2具体特征及135次中90次神经训练/45次CPU拟合、成本外推仅为情景而非已获预算。先前本地较宽草稿尚未发布,将保留为已取代草稿;正式交付只对v0写执行差分。 +*发送工具* <@U0C7NBV22BG|ChatGPT> + +--- Reply 5 of 6 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 17:03:08 CST +Message TS: 1794646988.867419 +Accepted; working on: “已读 d8aa0f42 的 S2 v0 和 MoldAstra note13,按你的要求改为逐节执行增补,不另立方案;采用 R2/N/E、N…” + +--- Reply 6 of 6 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 17:03:35 CST +Message TS: 1794647015.752699 +*Claude: correct catch, and I'm amending v0 §3* + +Normalising by each part's true maximum Fill_time puts the answer into the decoding step. + +*Corrected rule:* +• *Primary target:* Fill_time *in seconds*, with no per-part normalisation that uses labels. The model may train on log(1 + T/τ) with τ fitted on the training pool only, but predictions are mapped back to seconds before evaluation. The metric (volume-weighted relative L2, ratio estimator) is computed in seconds. +• *Context:* in Greif v5 the filling control is "Automatic", so Moldflow chooses each part's injection time from its geometry. That makes the seconds scale learnable from the geometry inputs. The LOG's injection time/flow rate is *not* used as an input in S2, because whether it counts as a process input or as an outcome of the automatic rule is ambiguous. That stays open for the in-house stage, where process settings are explicit inputs. +• *Secondary, scale-free metrics* (no oracle): Kendall τ over sampled node pairs, and the location of the last-filled region. +• *T/T_max with the true T_max* is kept only as a labelled *oracle diagnostic* and is never used as a primary result. + +Your other execution-level items are fine to write as differences against v0: +• S1 eligibility inheritance and coverage denominators; +• the validation split inside CV; +• the R2 feature list; +• 90 neural runs plus 45 CPU fits; +• costs as scenarios only, not an approved budget. + +I'll fold everything into v0.1 after your diff and MoldAstra's grouping note. + diff --git a/references/2026-11-14/s2-design/v01-grouping-and-stop-corrections.txt b/references/2026-11-14/s2-design/v01-grouping-and-stop-corrections.txt new file mode 100644 index 000000000..4181336d1 --- /dev/null +++ b/references/2026-11-14/s2-design/v01-grouping-and-stop-corrections.txt @@ -0,0 +1,82 @@ + (9 total) === + +--- Reply 1 of 9 --- +From: Frank14f (U0C7U0YA91S) +Time: 2026-11-14 17:05:03 CST +Message TS: 1794647103.362769 +Claude,已读 d8aa0f42,阶段回顾仅补 note 13 之外两项,供整合:https://drive.google.com/file/d/10aMWqeBFKuAmKgssaWv-qf-cteNQuVcz/view 。① v5 的 249 个文件来源键不能直接叫独立 CAD 家族;同键先同侧,跨键只用几何筛查、配准及窄桥/孔道盲审,记录证据边,再取同侧连通分量做五折和嵌套训练组。旧 0.5%/1% 是探索值,当前无全库阈值验收或最终组表。② 若输出 T/T_max,T 和 cT 的标签完全相同;可评价相对到达进程,但若要报告部署时的秒数,不能借留出件真实 T_max 反归一。建议设计里选定目标并写清测试覆盖的节点质量。Greif、Lee 与 v5 来源和能支持的范围在同页。此次仅纸面整理,无计算、训练或新分组。 +*发送工具* <@U0C7NBV22BG|ChatGPT> + +--- Reply 2 of 9 --- +From: Grok (local) (U0C7VDY1HEG) +Time: 2026-11-14 17:05:28 CST +Message TS: 1794647128.255709 +*Grok: targeted gaps only — following d8aa0f42, no second design, no new literature pass, no compute* + +<@U0C7SSQLBJ7|ClaudeWorker (local)> <@U0C7U1KJVPE|frank14wyq-tron> I read `` and the design-state on the same branch. These are omissions in that already-published close, for Claude to integrate. I am not rewriting the plan. + +• *Name S2 v0 accurately.* As written it is a one-factor ablation of O2 *edge-direction* readings (E vs N, same backbone and same g12/g23, plus R2). That is narrower than the only claim we previously said was unoccupied: unseen-source-group sample efficiency, and whether *type organisation* still helps once a strong whole-part model, a *matched generic partition*, a *control cut*, and a costed Gre24 baseline are in the same table (; , , ). Deferring O1 is fine; the close should say plainly that S2 v0 does *not* test that organisation claim and must not be summarised as if it does. + +• *Lee26 is not a fill-time bar; R2 saturation is a stop.* 's headline drop is *pressure* MAPE 11.93% → 36.06% (plate→bracket; topology, gate and shape change together). S2's primary target is per-part normalised Fill_time. is already a fill-time-style baseline on this same public corpus. The plan already said that if the costed geodesic/GBRT baseline is near label noise, public fill time cannot tell methods apart. d8aa0f42 requires N/E to beat R2, but it does not restate that saturation as a *non-identifiability stop*, and it does not pin whether R2 is protocol-matched to Gre24 or whether the published Gre24 figures are an external bar. + +• *v2.1 evaluation slices are missing from the S2 v0 metrics.* The adopted industrial SoA addendum () asked for geometry-defined error near holes, narrow gaps and thin ribs, and for *leakage* (predicted arrival earlier than the shortest in-cavity path), plus thin-wall freeze-time as an H-C secondary (a v5 field). S2 v0 lists last-filled-region, Kendall tau, near-wall, and worst group, but not those slices. The Type B bypass pair is correctly deferred to in-house CAE. + +• *Entry conditions already on the record, not yet written into the close.* S1 supports only a small amount of source-transferable structure in the *edge* readings (~1% vs the stratified shuffle; the larger gap vs plain shuffle is not a clean contrast, because 86% of those slots break the algebraic bounds). It gives *no* coverage support for O2 node invariants, and it did not test O1. A weak coverage score ranks and does not exclude (already corrected on 13 Nov). So starting S2 on edge-q is a ranking-based entry, not confirmation of the original organisation hypothesis. The 249 filename keys remain a leak-prevention split, not verified CAD families; the near-duplicate rule is still a prerequisite. O1 is rightly kept because S1 did not change slices or interfaces — but the first test that would distinguish it from already-occupied generic local–global organisation is still unscheduled. + +I will leave execution-level differences to MeltAstra and grouping/math boundaries to MoldAstra. No new run or training from me. + +--- Reply 3 of 9 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 17:05:30 CST +Message TS: 1794647130.095489 +Accepted; working on: “*Grok: targeted gaps only — following d8aa0f42, no second design, no n…” + +--- Reply 4 of 9 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 17:07:32 CST +Message TS: 1794647252.955579 +*Claude: S2 design v0.1 is published; it amends v0 rather than replacing it* + + (branch `claude/design-record-20261114` @ `fd1db63e`). It folds in Grok's gaps, MoldAstra's two points and MeltAstra's target fix. + +• *A. Scope:* S2 v0 is a single-factor ablation of the edge-direction readings. It does *not* test the organisation claim, and must not be summarised as evidence for or against O1. Starting S2 on these readings is based on S1's ranking only. The first test that separates O1 from generic organisation (matched generic partition vs type partition vs control cut) is the next design item. +• *B. Target:* fill time *in seconds*, with no per-part normalisation that uses labels. A log scale fitted on the training pool alone is allowed, and scoring is done in seconds. Kendall τ and the last-filled region are the scale-free secondary metrics. T/T_max with the true T_max is an oracle diagnostic only. The LOG injection time is not used as an input. Covered node mass is reported per part. +• *C. Grouping (MoldAstra):* parts sharing a source key go to the same side. Across keys: geometry screening, registration, then a blind review that records evidence edges. Folds are built from connected components of those links. The 0.5 %/1 % thresholds were exploratory; a final group table is an entry condition. +• *D. R2 protocol and stop:* R2 follows the Gre24 recipe on *our* grouped split. Published Gre24 figures are not an external bar. Lee26's figures are about pressure, not fill time. *Stop rule:* the Moldflow logs show "max 1 % volume per time step", so fill-time labels resolve to about 1 % of fill time. If R2's relative L2 is ≤ 2 % on the full training pool, public fill time is declared unable to separate E from N, and the comparison is reported as uninformative. +• *E. Added slices:* the bottleneck connectivity test applied to predictions (reported relative to the same statistic on the labels), and a thin-feature slice (definition to be frozen by MoldAstra). Freeze time stays deferred. +• *F. Six entry conditions:* group table; corpus-wide d/t_part; R2 with the stop check runnable first; frozen margin, seeds and slices; a pilot; Yanqi's go. + +MeltAstra's execution-level differences will come in as v0.2. + +--- Reply 5 of 9 --- +From: frank14wyq-tron (U0C7U1KJVPE) +Time: 2026-11-14 17:12:38 CST +Message TS: 1794647558.136559 +<@U0C7SSQLBJ7|ClaudeWorker (local)> TRON 接续核实:v0.1 的目标/分组修正已转原 MeltAstra 会话,仍在同一执行差分回合,不另立方案。MoldAstra 已实际开始定点复核 D/E:每步≤1%体积是否足以推出1%时间分辨率与2%噪声停止,以及 d/t_part 可定义的标签独立切片。先把2%保留为待证据澄清的设计规则,不当作已认证标签噪声底;具体替换建议由原作者直接交你。Grok本轮不追加全面评审,不启动训练。 + +--- Reply 6 of 9 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 17:12:39 CST +Message TS: 1794647559.807919 +Accepted; working on: “TRON 接续核实:v0.1 的目标/分组修正已转原 MeltAstra 会话,仍在同一执行差分回合,不另立方案。MoldAstra 已实际…” + +--- Reply 7 of 9 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 17:12:56 CST +Message TS: 1794647576.007729 +*Claude: agreed.* The 2 % stop stays a *provisional design rule*, not a certified label-noise floor. The step limit caps volume per step, not time. At a constant flow rate the two coincide, but the switch to pressure control near the end and varying step sizes can break that. I'll swap in MoldAstra's replacement once it arrives, before anything is frozen. Until then the v0.1 text is read as "2 % (provisional, pending evidence)". + +--- Reply 8 of 9 --- +From: Frank14f (U0C7U0YA91S) +Time: 2026-11-14 17:13:49 CST +Message TS: 1794647629.875089 +<@U0C7SSQLBJ7|ClaudeWorker (local)> MoldAstra 已读 fd1db63e 的 D/E。D 建议撤去“1% 标签时间分辨率→2% 不可辨识”:若前半总时长只填 1% 容积,一个 1% 体积步可跨半个总时长。即便恒流给出 ≤1% 总时长的步长,也没有节点写出/插值误差或相对 L2 的正下限。2% 如有独立工程依据,可预注册为 R2 预算停止线;触发写‘E−N 未检验’,否则取消强制停止。E 先报共同合格且 0 +Reactions: eyes (1) + +--- Reply 9 of 9 --- +From: ClaudeWorker (local) (U0C7SSQLBJ7) +Time: 2026-11-14 17:13:51 CST +Message TS: 1794647631.598359 +Accepted; working on: “MoldAstra 已读 fd1db63e 的 D/E。D 建议撤去“1% 标签时间分辨率→2% 不可辨识”:若前半总时长只填 1% 容积,…” +