Absorb S2 v0.3 rulings, verify pinned licenses, and maintain scope-specific entry status
This commit is contained in:
1 parent
33b18e0831
commit
3d3c2bd4e8
10 files changed
+459
-24
No files matched your search
@@ -16,4 +16,4 @@ Claude 统筹整合与主审;MeltAstra 提供执行、数据和预算证据;
|
||||
|
||||
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 与 `4e435a03` 的 v0.1a 及 `72deb266` 的 v0.2 裁量和 `3ae89805` 的阶段回顾为共同基线,不另建第二套方案。
|
||||
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 与 `4e435a03` 的 v0.1a 及 `72deb266` 的 v0.2 和 `69c58a77` 的 v0.3 裁量和 `3ae89805` 的阶段回顾为共同基线,不另建第二套方案。
|
||||
@@ -1,19 +1,19 @@
|
||||
# S2 执行差分:v0.2 裁量对照与剩余定义
|
||||
# S2 执行差分:v0.3 裁量对照与入口状态
|
||||
|
||||
**当前统一设计:** Claude [v0.2 / 72deb2664411d0256895108b7ab108e2562ebf0b](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/72deb2664411d0256895108b7ab108e2562ebf0b/docs/claude/s2-design-v0.2-2026-11-14.md),[17:20:04主审回执](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794648004314569)。本文件由 MeltAstra 实读全文后更新,仅对照已经作出的裁量和剩余定义,不另建方案。
|
||||
**前序统一设计:** Claude [v0.2 / 72deb2664411d0256895108b7ab108e2562ebf0b](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/72deb2664411d0256895108b7ab108e2562ebf0b/docs/claude/s2-design-v0.2-2026-11-14.md),[17:20:04主审回执](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794648004314569)。**当前裁量为 [v0.3 / 69c58a77006325319903f7862054d461b00ca75c](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/69c58a77006325319903f7862054d461b00ca75c/docs/claude/s2-design-v0.3-2026-11-14.md),[17:28:01回执](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794648481660359)。** MeltAstra已完整实读并在原文件接续,不另建方案。前轮对照稿固定于[33b18e08](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/33b18e0831ecbc86586f6d0c5d512dbb7adfd4b9/docs/s2-execution-diff-v0-2026-11-14.md)。
|
||||
|
||||
Claude 声明此前执行差分除明确改判项外均被采纳。先前提案与完整论证仍保存于 [c35b02d9 固定版本](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/c35b02d9e3b6814447d00c9b8950e29237dbb127/docs/s2-execution-diff-v0-2026-11-14.md),不删除或覆盖历史证据。v0 d8aa0f42 → v0.1 fd1db63e → v0.1a 4e435a03 → v0.2 72deb266,后版明确改判处优先。**本轮仍只有纸面整理,无预处理、pilot、拟合或训练;文稿中的资源数字是计划上限,不是执行许可。**
|
||||
Claude 声明此前执行差分除明确改判项外均被采纳。先前提案与完整论证仍保存于 [c35b02d9 固定版本](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/c35b02d9e3b6814447d00c9b8950e29237dbb127/docs/s2-execution-diff-v0-2026-11-14.md),不删除或覆盖历史证据。v0 d8aa0f42 → v0.1 fd1db63e → v0.1a 4e435a03 → v0.2 72deb266 → v0.3 69c58a77,后版明确改判处优先。**本轮仍只有纸面整理,无预处理、pilot、拟合或训练;文稿中的资源数字是计划上限,不是执行许可。**
|
||||
|
||||
## 1. 已吸纳与改判登记
|
||||
|
||||
| 原差分 | v0.2裁量 | 当前状态 |
|
||||
| 原差分 | v0.2及v0.3裁量 | 当前状态 |
|
||||
|---|---|---|
|
||||
| D1 目标 | R1:秒单位评分;若启用log变换,τ按每个外折、每档学习池,仅由该池训练标签拟合;大池不能给小池供尺度 | **已吸纳**。不再询问尺度可用数据范围;测试真实Tmax和LOG工艺输出禁入沿用 |
|
||||
| D1 目标 | R1/G2:秒单位评分;log变换已启用,τ是每个外折/学习池合格draw标签按次数加权的中位数;大池不能给小池供尺度 | **已吸纳**。不再询问尺度可用数据范围;测试真实Tmax和LOG工艺输出禁入沿用 |
|
||||
| D2 来源/拆分 | R2:来源键∪确认近重复证据边的连通分量;可能改组的未决对使S2 HOLD。5外折,seed 2026111409;非测试组件固定20%内验证;其余池取25/50/100%嵌套前缀 | **已吸纳**。20%不再是建议;development_exposed保留原五折并另报暴露/未暴露子集,无另设untouched集合 |
|
||||
| D3 资格/测度 | R3:d>0、谱有效、N_eff≥64、边块完整且有限;所有臂共用;**删除S1的gate q资格** | **已裁定**。需版本化新资格,不把S1资格位图当S2资格;此处去除的是模型不使用的资格要求,原始门信息和证据仍保留共享 |
|
||||
| D3 抽样/评分 | 24件复用冻结draw;新增件同规则seed [2026111400,prefix],每件固定2,000槽、不逐epoch刷新;件比率→组件内等件→组件等权,原来源键加权仅描述 | **已吸纳**。覆盖、NA和拒绝原因逐件报告;不能称全件预测结果 |
|
||||
| D4 R2 | R5:**R2-G,Gre24-style approximation**;既定9列、真实tet边Dijkstra,HistGradientBoostingRegressor + sample_weight,最多12个HPO配置;固定random_state | **已裁定**。不是Gre24复现;**15个外折×池最终拟合任务**,HPO额外计费;原45次CPU计划已被取代 |
|
||||
| D5 主干 | R5:Transolver主选、GINO备用,全节点非O2输入encoder;查询节点MLP decoder,N有g12/g23、E再加4列edge q;128/256宽、4层两档,同一档用于N/E | **已裁定设计**。源码commit/许可证仍须在pilot前实核,不能据设计中的MIT字样声称已完成核验;不再另选第三种主干 |
|
||||
| D5 主干 | R5:Transolver主选、GINO备用,全节点非O2输入encoder;查询节点MLP decoder,N有g12/g23、E再加4列edge q;128/256宽、4层两档,同一档用于N/E | **已裁定设计**。v0.3已给确切commit;本轮已只读回核两份MIT LICENSE(§2.1),未核验模型运行兼容性;不再另选第三种主干 |
|
||||
| D6 统计 | R6:100%池、整个共同主域,唯一主对比Δ=1−Ē_E/Ē_N;3匹配种子,m=10%;组件bootstrap 2,000次、seed2026111408、双侧95%百分位区间;保留成对种子块 | **已冻结主口径**。25/50%及切片等为次要,无多重性裁决;N对R2区间下界>0才称神经路线有用;查询MC误差不混入 |
|
||||
| D7 资源 | R7:10件分层预处理测量;每件30分钟/1.25GiB,不换件;pilot一外折100%池一种子、N/E各两容量档,≤4 runs、总≤2 GPU-h、单V100 16GB、AMP及FP32损失/归一;封存测试折 | **计划已具体化,执行仍未获go**。总账90神经+15 R2最终拟合+HPO+pilot+重跑;不能继续用“135总运行” |
|
||||
| D8 连通/薄区 | R4:全图瓶颈诊断延期,禁止真值填补、查询诱导图和O2零填。薄区沿v0.1a,掩码未验证前只报近壁与unknown | **缺口已通过延期处理**。不再要求当前S2补全图O2,不追加该计算预算 |
|
||||
@@ -25,22 +25,55 @@ Claude 声明此前执行差分除明确改判项外均被采纳。先前提案
|
||||
|
||||
主域是d>0抽样中的共同合格域。逐件保留正距质量份额h、2,000槽、不同ID、K、K/2000、h·K/2000、近壁份额、边界质量1−h和拒绝原因。删除gate q资格可能改变S2的K,不能沿抄S1覆盖值。末到区域与Kendall只命名为查询覆盖内的指标。
|
||||
|
||||
100%主对比的组件均值再跨三个匹配种子平均;bootstrap对两臂复用组件索引再计算比值。v0.2已选择组件重采样,不再沿用旧稿额外重采样种子这一候选;实现约定为抽到某组件时保留其三个完整匹配种子及对应外折预测。区间条件于已有三次训练实现、固定折与相互重叠的训练集,不声称已覆盖重新抽取训练集/种子的全部不确定性。
|
||||
100%主对比的组件均值再跨三个匹配种子平均;bootstrap对两臂复用组件索引再计算比值。v0.3已明确仅组件重采样,不再沿用旧稿额外重采样种子这一候选;已吸纳约定为抽到某组件时保留其三个完整匹配种子及对应外折预测。区间条件于已有三次训练实现、固定折与相互重叠的训练集,不声称已覆盖重新抽取训练集/种子的全部不确定性。
|
||||
|
||||
R2按每个外折/池的同一预测配对三个神经种子,不伪造三个独立R2随机实验。90与15是完成容量选择后的主矩阵计数,容量搜索、HPO、pilot、失败/重跑另列。
|
||||
|
||||
## 2. 剩余定义缺口(不是重开已裁定问题)
|
||||
## 2. G2–G5已裁量内容与G6入口表
|
||||
|
||||
| 编号 | 尚需写实的定义/产物 | 归属与边界 |
|
||||
### 2.1 已吸纳,不重开
|
||||
|
||||
| 原项 | v0.3明确内容 | 核验与执行状态 |
|
||||
|---|---|---|
|
||||
| G1 分组定标 | 物理尺度、配准/刚体退化、候选阈值、镜像/缩放模板同侧规则、窄桥孔道审边、筛外抽查和未决处理;确定离散分组/嵌套前缀的排序、取整和并列规则;完整development_exposed件号表 | MoldAstra已有note14框架、Claude冻结;**本轮不再生成候选、配准或分配实际fold**。249是文件键,不是已证实CAD家族 |
|
||||
| G2 抽样/输入 | 指向已冻结PPS实现及prefix解析/节点顺序,映射谱有效与边块完整的具体判据;资格删除gate q后逐项对照;τ变换是否启用及正尺度计算公式;断开组件、门最近节点并列和图距离不可达的有限表示规则 | MeltAstra准备规格、Claude裁量;现有S1实现可作为来源,不凭空默认;全节点encoder的基础输入也须有明确有限值规则。无新抽样或O2计算 |
|
||||
| G3 代码/训练 | Transolver/备用GINO确切commit及许可证;slice/latent等完整配置、查询embedding读取、N/E投影参数差;优化器/损失/更新数/批次与尺寸处理;同容量选择的统一验证评分/并列规则 | 主选与两容量档不重开。读代码与写规格可继续;新实现/执行另按授权范围。不得依据测试折换主干/容量 |
|
||||
| G4 R2/辅助统计 | sklearn版本、固定random_state、≤12配置的实际网格、早停策略(显式说明禁用或如何使用组隔离验证,不依赖隐含默认);三个训练seed的整数表;Kendall点对数量/seed/并列、末到区域定义与位置度量、NA组件对成对比较的处理 | Grok只执行Claude已提出的**Gre24精确特征/eikonal设定**定点核,不新增全文复审。N/E及R2用共同组/测度;最终15 fits不含最多15×12次HPO尝试的成本 |
|
||||
| G5 资源实施 | 10件无标签分层标准和预定件号;每件30分钟/RSS限额是整条准备流水线还是分步限额、子进程RSS统计;总CPU墙钟/并发/磁盘上限;pilot热身/更新/验证/存盘计时、主机RAM/磁盘/重试与最大件超限规则 | 已给出的上限照录,不放宽。明确硬件卡号与全部限额后仍须授权;失败保留且不换件/静默缩配。全库GPU-h仍等pilot实测后才可报 |
|
||||
| G6 阶段入口 | 逐项登记“定义冻结/实现审核/可执行产物/已运行回执”状态;全体阶段回顾及用户明确go | **不能将entry列表的顺序解释为先执行10件和pilot再补授权**。预处理测量、拟合和pilot各自也要落在用户明确批准的执行范围内;目前继续纸面设计 |
|
||||
| G2 抽样/资格 | 94011fd冻结draw;prefix取数字前缀整数、原tet节点ID升序。谱有效=O2-rep v1.1a valid;边块为所有真实入射边端均评估、至少一个channel-valid有限端、摘要有限 | 已只读核94011fd的draw contract及5e4b406驱动第97–106行。旧驱动eligible含gate判据,**不得直接复用该位图**;原始ID/次数重新应用S2资格,不重抽24件draw。未运行 |
|
||||
| G2 变换/距离 | y=log(1+T/τ),τ为当前学习池合格draw标签按次数加权中位数;图距源为YML门最近节点、等距取较小ID | 已吸纳。训练池之外的标签不参与τ;图距不可达或encoder输入非有限即整件失败留因,不插补。原624件单连通是既有审计记录,本轮未重算。空池/非正τ不能默认epsilon开跑 |
|
||||
| G3 模型/优化 | 4层、8heads、32slices、宽128/256;query decoder两个同宽隐藏层。AdamW lr1e−3、cosine、weight_decay1e−5、batch一件;y上次数加权MSE,件/组件等权;≤200epochs、内验证patience20 | 已吸纳。N/E同档;验证分数并列取小宽。只剩跨臂容量选取规则及额外搜索记账见A1,不重议层数、学习率或容量候选 |
|
||||
| G4 R2 | HistGBRT random_state=0,early_stopping=False;网格learning_rate{0.05,0.1}×max_leaf_nodes{31,63}×max_iter{200,500,1000},min_samples_leaf20;内验证选择 | 12配置已明确;sklearn版本在实现时固定,属于实现产物而非新增设计问题。15最终fit与HPO分账。Gre24近似命名不变,Grok只做原特征/eikonal定点核 |
|
||||
| G4 随机/辅助 | 神经seed{0,1,2}统一torch/numpy/data-order;Kendall每件5,000对不同合格ID、SeedSequence([2026111410,prefix]);末到位置为查询argmax预测与真值的欧氏距/Rg;对比仅双方有效组件、报告排除数 | 已吸纳数值、域和配对原则。Kendall真值并列剔除与tau-b名称冲突、argmax并列和组件有效聚合等仅余A2/A4的实施定义 |
|
||||
| G5 测量 | 按节点数三分位各3件、seed2026111411,加最大件00000666;每件整条准备流水线≤30min、整个进程树RSS≤1.25GiB;顺序执行,总≤3h、磁盘≤5GiB;失败不换件 | 上限均已定,不再列待裁量;唯一抽样重叠问题见A3。总3h是硬停,不承诺10件必完成;余件登记未执行,不能超时补足 |
|
||||
| G5 pilot | N/E×两容量,≤4runs,总≤2GPU-h;每run≤30min、VRAM≤15GiB、host RAM≤16GiB、0重试;一V10016GB、AMP/FP32损失与归一、测试折封存 | 均为待授权计划。准备/加载/热身/验证/存盘纳入墙钟监督,不能只计优化步骤;监控进程树、显存和写盘的监督器尚未实现/核验。不能靠OOM重跑缩配 |
|
||||
|
||||
G1–G5是统一设计落地所需定义和产物,并非要求每项再作一次全面研究。薄区的几何代理和容差沿MoldAstra note15,未验证则按已有裁量仅报近壁;全图瓶颈指标已延期,不列成当前S2的阻塞项。2%噪声底/强制停止已在v0.1a撤销,不重新引入。训练go不能由信息分享授权或“评审通过”替代。
|
||||
**源码/许可核验边界。** 已直接回读下列固定revision的LICENSE,均写明MIT;此核验只覆盖许可证原文及可寻址revision,不代表选定模型文件/API/依赖、全件query编码适配、V100数值/内存均已验证。后续改动仍须保留diff并标为改编。
|
||||
|
||||
- [thuml/Transolver@75e0f67643806a81cd1d3f6adc88dd8c02416fe7/LICENSE](https://raw.githubusercontent.com/thuml/Transolver/75e0f67643806a81cd1d3f6adc88dd8c02416fe7/LICENSE):1,084 bytes,SHA256 `2c919cd03fa823bf7eefc00a957ff8324c865cd22aa5285e563dc4b558084f25`,匹配v0.3给出的前缀。
|
||||
- [neuraloperator@a6648eeeb60fe0837f2d7d021517a945a159a0c8/LICENSE](https://raw.githubusercontent.com/neuraloperator/neuraloperator/a6648eeeb60fe0837f2d7d021517a945a159a0c8/LICENSE):1,082 bytes,SHA256 `5f335b0603acfb1d180bf3117492e2076562075cc9b6353a3b94d5d5bc03ef33`。
|
||||
|
||||
原文、URL和完整哈希保存于本包 `v03-upstream-license-readback.json`。同时保留冻结draw contract及原驱动源码快照;没有执行这些脚本。
|
||||
|
||||
### 2.2 一次集中交Claude的剩余实现歧义
|
||||
|
||||
以下四项会导致不同选模、指标或样本,不是将已定参数重新列待定。推荐写法仍是**提议,尚待Claude确认**;不据此运行。
|
||||
|
||||
1. **A1 共同容量的跨臂选择和适用范围。** 两臂各有内验证分数,v0.3的“最低验证误差”尚未指定如何变成一个共同宽度。建议同一外折/池按N/E匹配验证误差均值、用预定选择seed0选宽(并列小宽),之后同宽跑三个seed。请Claude明确是这个范围,还是pilot一次选宽后全矩阵固定;若每外折/池都试两宽,额外选型训练必须另计,不能仍将90次当全部训练费用,也不能把某外折的验证选择悄悄用于其他折测试后仍称无暴露。
|
||||
2. **A2 Kendall名称与并列处理。** v0.3同时写“tau-b”和“真值并列对排除”。标准tau-b的分母含仅真值并列及仅预测并列对的计数,预先剔除真值并列对会改变统计量。[SciPy官方定义](https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.kendalltau.html)。建议保留真值/预测并列计数,报告5,000对上的tau-b估计(非全体ID精确tau-b),不报由其直接推导的显著性p;或者保留排除规则但改名为条件成对排序分数并写分母。点对建议从不同合格ID均匀抽取,每对内部不同、对与对可重复,PCG64配既定SeedSequence;不足2个ID或零分母为NA。
|
||||
3. **A3 十件抽样保证唯一。** 最大件可能也在最高三分位抽中,照字面“3+3+3再加最大件”可能只有9个不同件。建议先按全库节点数排序分三层(同数按numeric prefix/完整case),从最高层的随机候选中先排除00000666,再每层无放回取3件,最后固定加入00000666,名单在运行前封存。这是预先选样规则,不是失败后换件;本轮未生成名单。
|
||||
4. **A4 末到并列与组件NA。** 建议预测/真值argmax并列各取最小原ID;一个组件在某对比中只有在两臂全部预定部件、全部3匹配seed上均有效才纳入,避免不同臂删不同件后仍称等组件配对。报告被排除的组件/件/抽样质量和失败原因,并标为完成者子集;不得用NA排除掩盖数值失败。请确认此严格有效定义或给出另一固定共同掩码规则。τ采用累计次数首次达到一半的下加权中位数;空集、τ≤0或非有限时报错/HOLD,不默认插补。
|
||||
|
||||
G1继续由MoldAstra提供纸面定标规则、Claude冻结;无候选重生成、配准或实际分组。薄区沿note15未验证前只报近壁;全图瓶颈已延期,2%强制停止已撤销,二者不再列当前缺口。Grok不追加全面评审。
|
||||
|
||||
### 2.3 G6 入口状态表
|
||||
|
||||
| 入口 | 当前状态 | 下个可核验凭据/授权边界 |
|
||||
|---|---|---|
|
||||
| 分组规则与组件 | G1纸面定标由MoldAstra接续;最终组件/折/嵌套池未生成 | 规则冻结、证据边、未决对清单、映射与development_exposed完整表;相关新计算须有对应授权 |
|
||||
| 目标域、抽样、资格 | 设计已定,历史源码已只读核对;S2资格适配尚未实现/验收 | 原始draw身份核对+新资格版本;不会误用S1门方向过滤后的集合 |
|
||||
| 代码、训练配置及统计 | G2–G5已吸纳;LICENSE核验完成;模型/监督器未适配;A1–A4集中待裁量 | 固定配置、源码diff、软件版本、指标定义与纸面审查记录;文稿不能冒称运行验收 |
|
||||
| 10件预处理测量 | **未执行、未获本步骤go**;计划30min/1.25GiB每件,总3h/5GiB | 名单/监督契约先冻结,再获Yanqi该步骤go;运行后逐件CPU/墙钟/进程树RSS/磁盘与超限回执 |
|
||||
| pilot | **未执行、未获本步骤go**;4runs/2GPU-h等上限仅计划 | 本步骤单独go及监督器;实际每run计时/显存/主存/数值/停止回执,失败0重试 |
|
||||
| 全矩阵与总预算 | **未执行、未获本步骤go**;主矩阵90神经+15R2最终fits,另计HPO/选型/pilot/失败 | pilot实测外推完整成本、全体阶段回顾、Yanqi完整范围单独go,不能以计划时数代替实测 |
|
||||
| O1/N0/内部CAE | 独立开放设计,未运行 | 按R8与研究初衷讨论;信息共享授权不扩张为这些计算许可 |
|
||||
|
||||
进入列表只表示顺序。**10件测量、pilot、全矩阵各需Yanqi针对该步骤明确go**;全体汇总、回顾初衷与前人工作、充分讨论仍是推进习惯。本轮仅文稿、来源核验和发布。
|
||||
|
||||
## 3. 阶段回顾:为什么做这一步,仍未回答什么
|
||||
|
||||
@@ -52,10 +85,11 @@ G1–G5是统一设计落地所需定义和产物,并非要求每项再作一
|
||||
|
||||
**O1保持开放。** 类型、真实孔/通道和区域连接图分别定义;网格细分环不等于真实孔洞,旧三跳标签也不是已验证语义类型。遵循v0.2 R8:类型定义→890+一无缓存件校准→区域表→容量LP→验收;本轮不执行。后续类型切分/通用切分/控制切法与整件参考须同准入池、同容量/预算比较,不能拿较小可分块子集对更大池模型作类型归因。内部CAE保留定压与定流控制,当前仅数据需求/控制表,不求解。N0只有用户要求检验整体O2价值时才扩展。
|
||||
|
||||
**协作与进入决定。** Claude主审整合,MoldAstra负责数学/分组边界,MeltAstra负责执行规格与证据,Grok完成既定定点核查;沿用 `docs/collaboration-stage-review-policy.md`。重大结果与进入新阶段前全体汇总、回顾初衷/前人、讨论分歧。**当前状态:v0.2设计裁量已接收;执行定义与可执行产物未齐,无新运行回执,无训练go。**
|
||||
**协作与进入决定。** Claude主审整合,MoldAstra负责数学/分组边界,MeltAstra负责执行规格与证据,Grok完成既定定点核查;沿用 `docs/collaboration-stage-review-policy.md`。重大结果与进入新阶段前全体汇总、回顾初衷/前人、讨论分歧。**当前状态:v0.3裁量已接收;G1及四项实现歧义待闭合,可执行产物/监督器未验收,无新运行回执,各步骤均未获go。**
|
||||
|
||||
## 4. 来源与交付范围
|
||||
|
||||
- [Claude v0.3 / 69c58a77](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/69c58a77006325319903f7862054d461b00ca75c/docs/claude/s2-design-v0.3-2026-11-14.md)及[17:28原串消息](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794648481660359)。
|
||||
- [Claude v0.2 / 72deb266](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/72deb2664411d0256895108b7ab108e2562ebf0b/docs/claude/s2-design-v0.2-2026-11-14.md)及[原串主审消息](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794648004314569)。
|
||||
- [前版执行差分 / c35b02d9](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/c35b02d9e3b6814447d00c9b8950e29237dbb127/docs/s2-execution-diff-v0-2026-11-14.md),保留完整提案及历次依据。
|
||||
- [MoldAstra note14](https://drive.google.com/file/d/10aMWqeBFKuAmKgssaWv-qf-cteNQuVcz/view)、[note15](https://drive.google.com/file/d/18X07uNtM1j697MCztXryzmITCcud3_T4/view)、[v2.1六区互斥表](https://drive.google.com/file/d/1EdAhFr6gn3R-hlLDqqpRYD8--1mWtZBM/view);本轮沿用已经取得的来源快照。
|
||||
|
||||
@@ -0,0 +1,54 @@
|
||||
# S2 design v0.3: rulings on remaining gaps G2–G5 (Claude, 2026-11-14)
|
||||
|
||||
Integrates MeltAstra's diff at 33b18e08, §2. G1 (grouping calibration) stays with MoldAstra and is frozen by me later. G6 is a status table kept by MeltAstra.
|
||||
|
||||
No execution of any kind (preprocessing measurement, pilot, fits or training) without Yanqi's explicit approval for that scope. The entry list is an order, not an authorisation.
|
||||
|
||||
## Clarifications (MoldAstra, accepted)
|
||||
- **Bootstrap:** each replicate resamples components only. The three matched seed runs of each component travel with it, and seeds are not resampled. Both arms are recomputed on the same indices, giving Δ* = 1 − Ē_E*/Ē_N*. The interval is conditional on the fixed folds and the trained runs; it does not cover retraining variance.
|
||||
- **Reused S1 draws:** S2 eligibility is re-applied from the raw IDs and counts, never from S1's already-filtered sets.
|
||||
|
||||
## G2: sampling and inputs
|
||||
- **Draw:** the frozen implementation of commit 94011fd. prefix = integer numeric prefix. Candidate order = ascending original tet-node ID.
|
||||
- **Eligibility:**
|
||||
- valid spectrum = O2-rep v1.1a status valid;
|
||||
- edge block complete = every incident true-edge end evaluated, at least one channel-valid finite end, and finite summaries. This is the definition used by driver 5e4b406.
|
||||
- **Target transform:** enabled.
|
||||
- τ = draw-count-weighted median of T over the eligible draws of the learning pool (per fold and pool).
|
||||
- y = log(1 + T/τ); predictions are inverted before scoring.
|
||||
- **Graph distance:** ℓ_g is Dijkstra on tet edges from the node nearest the YML gate; ties go to the smaller ID.
|
||||
- All 624 tet subdomains are single components (audit 2026-11-13), so "unreachable" should not occur.
|
||||
- If it does, or if any encoder input is non-finite, the part fails closed with a reason. No imputation.
|
||||
|
||||
## G3: code and training
|
||||
- **Code:**
|
||||
- primary: thuml/Transolver @ 75e0f67643806a81cd1d3f6adc88dd8c02416fe7 (MIT, LICENSE sha256 2c919cd0…);
|
||||
- backup: neuraloperator @ a6648eeeb60fe0837f2d7d021517a945a159a0c8 (MIT).
|
||||
- Any modifications are recorded as a diff and never called a reproduction.
|
||||
- **Configuration:**
|
||||
- 4 layers, 8 heads, 32 slices, width ∈ {128, 256};
|
||||
- query decoder MLP with 2 hidden layers of the same width;
|
||||
- N decoder input = [embedding, g12, g23]; E adds the 4 edge q values. Parameter counts are reported.
|
||||
- **Optimisation:**
|
||||
- AdamW, lr 1e-3, cosine schedule, weight decay 1e-5, batch = 1 part;
|
||||
- loss = draw-count-weighted MSE in y, normalised so that parts within a component and components overall are equally weighted;
|
||||
- at most 200 epochs, early stopping on inner validation with patience 20.
|
||||
- **Capacity choice:** lowest inner-validation component-equal error; a tie goes to the smaller width. The chosen width is applied to both N and E.
|
||||
|
||||
## G4: R2 and auxiliary statistics
|
||||
- **R2:** sklearn HistGradientBoostingRegressor, version pinned at implementation, random_state=0, early_stopping=False (its internal split is not group-aware).
|
||||
- Grid (12 configurations): learning_rate {0.05, 0.1} x max_leaf_nodes {31, 63} x max_iter {200, 500, 1000}; min_samples_leaf = 20.
|
||||
- Selected on inner validation.
|
||||
- **Neural seeds:** {0, 1, 2}, each applied to the torch, numpy and data-order RNGs.
|
||||
- **Kendall tau-b:** 5,000 random pairs of distinct eligible IDs per part, seed SeedSequence([2026111410, prefix]). Pairs tied in the truth are excluded.
|
||||
- **Last-filled location:** among eligible query nodes, the distance between argmax T̂ and argmax T, divided by R_g. Labelled as being within query coverage only.
|
||||
- **NA handling:** a paired contrast uses components where both arms are valid. The excluded count is reported.
|
||||
|
||||
## G5: resources (upper limits; still require approval)
|
||||
- **10-part measurement:** tertiles by node count, 3 parts per tertile drawn with seed 2026111411, plus the corpus's largest part (00000666).
|
||||
- Per part, the whole preprocessing pipeline: 30 min wall time, 1.25 GiB RSS across the whole process tree.
|
||||
- Run sequentially; total ≤ 3 h; disk ≤ 5 GiB.
|
||||
- A failure is recorded and the part is not swapped.
|
||||
- **Pilot runs:** per run, wall time ≤ 30 min, VRAM ≤ 15 GiB, host RAM ≤ 16 GiB; 0 retries.
|
||||
- The overall cap stays at 4 runs and 2 GPU-h.
|
||||
- Test fold sealed.
|
||||
@@ -5,7 +5,8 @@
|
||||
"d8aa0f420647701709e03629326dce8339db739a",
|
||||
"fd1db63ee1c4fb4415a36ed95a2eedfb73a5c85c",
|
||||
"4e435a03d3e1023bb2c12154ab4afc98b37da741",
|
||||
"72deb2664411d0256895108b7ab108e2562ebf0b"
|
||||
"72deb2664411d0256895108b7ab108e2562ebf0b",
|
||||
"69c58a77006325319903f7862054d461b00ca75c"
|
||||
],
|
||||
"files": [
|
||||
{
|
||||
@@ -15,13 +16,13 @@
|
||||
},
|
||||
{
|
||||
"path": "docs/collaboration-stage-review-policy.md",
|
||||
"bytes": 2321,
|
||||
"sha256": "7757dc15066cda6e74a3b6aed559c6cf42a3a6e982e711456e63d71be41522ce"
|
||||
"bytes": 2345,
|
||||
"sha256": "7d33251cded22384e675b106f6e617dcc2a570b3ec543f61aea744accf591dd7"
|
||||
},
|
||||
{
|
||||
"path": "docs/s2-execution-diff-v0-2026-11-14.md",
|
||||
"bytes": 12048,
|
||||
"sha256": "91219858ffe7fedb71e371e9543dd85914b3a87adaf53702a89a703fc7d0aae6"
|
||||
"bytes": 18160,
|
||||
"sha256": "9f744bc93b473b944ce0447949d234f86c2505d949266dc93e4a6262bf13af45"
|
||||
},
|
||||
{
|
||||
"path": "docs/sharing-authorization-2026-11-14.md",
|
||||
@@ -73,11 +74,21 @@
|
||||
"bytes": 6892,
|
||||
"sha256": "bd59621b1507a12e24bf3221cb4481611ec1695c69d7f79c071996d3e175c0bb"
|
||||
},
|
||||
{
|
||||
"path": "references/2026-11-14/s2-design/claude-s2-v03-69c58a77.md",
|
||||
"bytes": 4307,
|
||||
"sha256": "f8b60afae6355ec429a08ccc06698c3d8b3cb73cb3058a851beb816746523d2f"
|
||||
},
|
||||
{
|
||||
"path": "references/2026-11-14/s2-design/claude-stage-review-3ae89805.md",
|
||||
"bytes": 4592,
|
||||
"sha256": "e7b7221ba5e5419a7ce3ee59a0d6f383210668ac39ddb74abf97362a1dc9e0da"
|
||||
},
|
||||
{
|
||||
"path": "references/2026-11-14/s2-design/frozen-draw-contract-94011fd.json",
|
||||
"bytes": 2511,
|
||||
"sha256": "c3631c5a1a48707014f820aeead23fd950cf6b3601dad8073d5f5c30d7eba495"
|
||||
},
|
||||
{
|
||||
"path": "references/2026-11-14/s2-design/moldastra-note13.md",
|
||||
"bytes": 4513,
|
||||
@@ -93,16 +104,31 @@
|
||||
"bytes": 4391,
|
||||
"sha256": "60821afaf277dc339c1018223768a6ec224e4769ffc8d12ee3de5101b240dfc7"
|
||||
},
|
||||
{
|
||||
"path": "references/2026-11-14/s2-design/neuraloperator-LICENSE-v03.txt",
|
||||
"bytes": 1082,
|
||||
"sha256": "5f335b0603acfb1d180bf3117492e2076562075cc9b6353a3b94d5d5bc03ef33"
|
||||
},
|
||||
{
|
||||
"path": "references/2026-11-14/s2-design/s1-review-and-topology-steering.txt",
|
||||
"bytes": 4665,
|
||||
"sha256": "f0e33c2def861a576e5168471344b4ebf37623f2be2c6af74def6d13a40e422c"
|
||||
},
|
||||
{
|
||||
"path": "references/2026-11-14/s2-design/sampled-driver-5e4b406.py",
|
||||
"bytes": 11228,
|
||||
"sha256": "66ab07068c6553a5d49d5124087d69c338e057919e2fe0c31124ba6d2d2a5d36"
|
||||
},
|
||||
{
|
||||
"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/transolver-LICENSE-v03.txt",
|
||||
"bytes": 1084,
|
||||
"sha256": "2c919cd03fa823bf7eefc00a957ff8324c865cd22aa5285e563dc4b558084f25"
|
||||
},
|
||||
{
|
||||
"path": "references/2026-11-14/s2-design/v01-grouping-and-stop-corrections.txt",
|
||||
"bytes": 10822,
|
||||
@@ -112,7 +138,17 @@
|
||||
"path": "references/2026-11-14/s2-design/v02-main-review-receipt.txt",
|
||||
"bytes": 3699,
|
||||
"sha256": "75de33048a55ab53967d27c800374e6bcba3eb67480eccef16a8126c1d176b63"
|
||||
},
|
||||
{
|
||||
"path": "references/2026-11-14/s2-design/v03-main-review-receipt.txt",
|
||||
"bytes": 3545,
|
||||
"sha256": "76dcedcd6cec75074f3cead8665cacedcceeff801808936a829dd70df2393718"
|
||||
},
|
||||
{
|
||||
"path": "references/2026-11-14/s2-design/v03-upstream-license-readback.json",
|
||||
"bytes": 3057,
|
||||
"sha256": "b5adf4a29f800203aeb4ed4bf223b37f6fcf2a6d930106f5f7fb65cdea1a1372"
|
||||
}
|
||||
],
|
||||
"preceding_execution_diff": "c35b02d9e3b6814447d00c9b8950e29237dbb127"
|
||||
"preceding_execution_diff": "33b18e0831ecbc86586f6d0c5d512dbb7adfd4b9"
|
||||
}
|
||||
@@ -0,0 +1,48 @@
|
||||
{
|
||||
"version": "M2 v1.2 sampled",
|
||||
"frozen_before_first_draw_utc": "2026-11-14T06:33:55.731281+00:00",
|
||||
"precheck_commit": "81f56afaca38304a4b5f48921927f97669ebbac9",
|
||||
"panel_sha256": "c594110b017ad8c62a542e4f634ce2879ccd1c30bd818429e73c0910a57a1e69",
|
||||
"count_contract_sha256": "21d11a0d975735ffb4cba355aea8abe0fdfe4f3ef7d7f586444c6ad8113be342",
|
||||
"code_sha256": "7d58fcea9c407f7a9a42ce592d0cedffbb34772a3f3d1930c6b03f3dcd730a81",
|
||||
"code_path": "tools/draw_s1_sampled.py",
|
||||
"document_sha256": {
|
||||
"docs/s1-m2-v1.2-draw-current-source-2026-11-14.md": "512b42e6d103c7645397bce184d39a586dfc29703393f3061849179e8ad824a5",
|
||||
"docs/s1-m2-v1.2-draw-go-2026-11-14.txt": "704b4919d65934c179bbf0e1bf943112197ad50c90a0e919685ddbbb103a07e4"
|
||||
},
|
||||
"numpy": "1.26.4",
|
||||
"python": "3.10.13",
|
||||
"candidates": "All original-pose nodes with d>0, ascending original tetra-submesh ID; raw VTU IDs strictly ascending in same order",
|
||||
"distribution": "iid with replacement, p=mass[candidates]/np.sum(mass[candidates], dtype=float64)",
|
||||
"algorithm": "Generator(PCG64(SeedSequence([2026111400, numeric_prefix]))).choice(candidate_ids, size=2000, replace=True, p=probabilities)",
|
||||
"draw_count": 2000,
|
||||
"primary_seed": 2026111400,
|
||||
"permutation_seeds": [
|
||||
2026111401,
|
||||
2026111402,
|
||||
2026111403,
|
||||
2026111404,
|
||||
2026111405
|
||||
],
|
||||
"bootstrap_seed": 2026111406,
|
||||
"permutations_executed": false,
|
||||
"bootstrap_executed": false,
|
||||
"identity": "panel_index, case, source_key, numeric_prefix preserved; one draw per part reused across all folds",
|
||||
"fold_roles": null,
|
||||
"fold_status": "Actual fold assignment not specified in current document; must freeze before scoring. Source isolation mandatory.",
|
||||
"exact_visits": "Python integer sum of saved neighbor_counts[i] over ALL distinct drawn IDs, once per ID, before O2 eligibility",
|
||||
"common_eligible_draws": null,
|
||||
"n_ref": null,
|
||||
"eligibility_status": "Unknown until O2; no prefilter except d>0",
|
||||
"supersedes": "Only historical v1.1 reference/query/rank/shortfall clauses; original panel and count files unchanged",
|
||||
"full_source_field": "All four-point source positions/masses and original nodal-gradient tensors, full-part R/t_part/R_g retained for future O2",
|
||||
"future_per_field_wall_seconds": 1800,
|
||||
"future_per_field_rss_bytes": 1342177280,
|
||||
"timing_scenario_seconds_per_visit": 7.3e-07,
|
||||
"timing_scenario_is_bound": false,
|
||||
"visit_alarm": 500000000,
|
||||
"visit_hard_cap": null,
|
||||
"O2_executed": false,
|
||||
"S1_scoring": false,
|
||||
"training": false
|
||||
}
|
||||
@@ -0,0 +1,21 @@
|
||||
MIT License
|
||||
|
||||
Copyright (c) 2023 NeuralOperator developers
|
||||
|
||||
Permission is hereby granted, free of charge, to any person obtaining a copy
|
||||
of this software and associated documentation files (the "Software"), to deal
|
||||
in the Software without restriction, including without limitation the rights
|
||||
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
||||
copies of the Software, and to permit persons to whom the Software is
|
||||
furnished to do so, subject to the following conditions:
|
||||
|
||||
The above copyright notice and this permission notice shall be included in all
|
||||
copies or substantial portions of the Software.
|
||||
|
||||
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
||||
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
||||
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
||||
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
||||
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
||||
SOFTWARE.
|
||||
@@ -0,0 +1,157 @@
|
||||
"""Frozen E6/O2 routines on a fixed subset of anchors; no S1 scoring."""
|
||||
import argparse
|
||||
import hashlib
|
||||
import json
|
||||
from pathlib import Path
|
||||
import resource
|
||||
import time
|
||||
import numpy as np
|
||||
import run_e6_real_kernel as e6
|
||||
import o2_rep as rep
|
||||
|
||||
|
||||
def sha(p):return hashlib.sha256(Path(p).read_bytes()).hexdigest()
|
||||
def read(p):return json.loads(Path(p).read_text())
|
||||
def save(p,v):Path(p).write_text(json.dumps(v,indent=2,allow_nan=False)+'\n')
|
||||
|
||||
|
||||
def bit_comparison(name, actual, expected):
|
||||
a=np.asarray(actual);b=np.asarray(expected)
|
||||
same=a.shape==b.shape and a.dtype==b.dtype and a.tobytes()==b.tobytes()
|
||||
row=dict(array=name,shape=list(a.shape),dtype=str(a.dtype),bitwise_equal=same)
|
||||
if not same and a.shape==b.shape and a.dtype==b.dtype:
|
||||
av=np.ascontiguousarray(a).reshape(-1).view(np.dtype((np.void,a.dtype.itemsize)))
|
||||
bv=np.ascontiguousarray(b).reshape(-1).view(np.dtype((np.void,b.dtype.itemsize)))
|
||||
bad=np.flatnonzero(av!=bv);row['different_elements']=len(bad)
|
||||
row['first_flat_indices']=bad[:20].tolist()
|
||||
if np.issubdtype(a.dtype,np.number):
|
||||
finite=np.isfinite(a)&np.isfinite(b)
|
||||
row['max_abs_finite_difference']=float(np.max(np.abs(a[finite]-b[finite]),initial=0))
|
||||
return row
|
||||
|
||||
|
||||
def run(out,phase,index):
|
||||
c=read(out/'execution_contract.json');s=c[phase][index];case=s['case'];pose=s['pose']
|
||||
dest=out/phase/s['job'];dest.mkdir(exist_ok=False);start=time.monotonic()
|
||||
save(dest/'started.json',dict(contract_sha256=sha(out/'execution_contract.json'),case=case,pose=pose,phase=phase))
|
||||
for p,h in c['code_sha256'].items():assert sha(p)==h,p
|
||||
for p,h in s['input_sha256'].items():assert sha(p)==h,p
|
||||
if phase=='panel':
|
||||
gate=read(out/'equivalence_gate.json');assert gate['status']=='PASS'
|
||||
for job,h in gate['result_sha256'].items():assert sha(out/'equivalence'/job/'result.json')==h
|
||||
with np.load(s['geometry'],allow_pickle=False) as z:
|
||||
x=z['points_raw'];tet=z['tetra'];mass=z['node_volume_raw'];edges=z['edges'];gate=z['gate_location_raw']
|
||||
raw_ids=z['raw_node_ids'] if 'raw_node_ids' in z else np.arange(len(x))
|
||||
d=np.load(s['distance'],allow_pickle=False)
|
||||
with np.load(s['anchor_file'],allow_pickle=False) as z:
|
||||
ids=z['unique_ids'];multiplicity=z['multiplicity']
|
||||
assert np.all(np.diff(ids)>0) and len(multiplicity)==len(ids)
|
||||
# Original full-part scales, before eligibility; carry L_med to rotated pose.
|
||||
lengths=np.linalg.norm(x[edges[:,1]]-x[edges[:,0]],axis=1)
|
||||
positive=lengths[np.isfinite(lengths)&(lengths>0)]
|
||||
L=float(np.median(positive)) if len(positive) else None
|
||||
if 'L_med_raw' in s:assert L==s['L_med_raw']
|
||||
center=np.sum(mass[:,None]*x,axis=0)/mass.sum()
|
||||
Rg=float(np.sqrt(np.sum(mass*np.sum((x-center)**2,axis=1))/mass.sum()))
|
||||
if not np.isfinite(Rg) or Rg<=0:raise ValueError('Invalid full-part R_g')
|
||||
if pose=='rotated':
|
||||
Q=np.asarray(c['rotation']);shift=np.asarray(c['translation']);x=x@Q.T+shift;gate=gate@Q.T+shift
|
||||
load_done=time.monotonic()
|
||||
progress=dict(stage='raw_field',anchors_done=0,anchors_total=len(ids),neighbor_visits=0)
|
||||
save(dest/'progress.json',progress)
|
||||
raw,raw_meta=e6.raw_field(x,tet,d);raw_done=time.monotonic()
|
||||
sx,sm,st=e6.samples(x,tet,mass,'tetra_four_point',raw);sources_done=time.monotonic()
|
||||
source_count=len(sx);real_tree=e6.cKDTree;progress['stage']='kernel'
|
||||
class ProgressTree:
|
||||
def __init__(self,*args,**kwargs):self.tree=real_tree(*args,**kwargs)
|
||||
def query_ball_point(self,*args,**kwargs):
|
||||
result=self.tree.query_ball_point(*args,**kwargs)
|
||||
progress['anchors_done']+=1;progress['neighbor_visits']+=len(result)
|
||||
if progress['anchors_done']%32==0:
|
||||
save(dest/'progress.json',dict(progress,elapsed_seconds=time.monotonic()-start))
|
||||
return result
|
||||
e6.cKDTree=ProgressTree
|
||||
try:a=e6.integrate(x[ids],sx,sm,st,s['radius'])
|
||||
finally:e6.cKDTree=real_tree
|
||||
del sx,sm,st,raw
|
||||
kernel_done=time.monotonic()
|
||||
if phase=='panel':assert progress['neighbor_visits']==s['exact_unique_visits']
|
||||
r=rep.readout(a['tensor'],a['kernel_mass'],d[ids]>0,a['effective_samples'])
|
||||
checks=rep.acceptance(r);limit=c['frozen_algebra_tolerance']
|
||||
if checks['valid_count']:
|
||||
assert checks['mu_sum_error']<=limit and checks['mu_min']>=-limit and checks['ordering_violation']<=limit
|
||||
assert checks['B_symmetry_error']<=limit and checks['B_min_eigenvalue']>=-limit and checks['reconstruction_max']<=limit
|
||||
gq=rep.contract_direction(r['B'],r['mu'],r['valid'],gate-x[ids],L,'gate')
|
||||
readout_done=time.monotonic()
|
||||
lookup=np.full(len(x),-1,np.int64);lookup[ids]=np.arange(len(ids))
|
||||
edge_idx,side=np.nonzero(lookup[edges]>=0);anchor_local=lookup[edges[edge_idx,side]]
|
||||
m=len(edge_idx);q=np.empty((m,2));channel_valid=np.empty(m,bool);direction_valid=np.empty(m,bool)
|
||||
edge_length=np.empty(m);edge_reason=np.empty(m,np.uint8)
|
||||
for begin in range(0,m,4096):
|
||||
stop=min(begin+4096,m);e=edges[edge_idx[begin:stop]];j=anchor_local[begin:stop]
|
||||
vectors=x[e[:,1]]-x[e[:,0]]
|
||||
eq=rep.contract_direction(r['B'][j],r['mu'][j],r['valid'][j],vectors,L,'edge')
|
||||
q[begin:stop]=eq['q'];channel_valid[begin:stop]=eq['channel_valid']
|
||||
direction_valid[begin:stop]=eq['direction_valid'];edge_length[begin:stop]=eq['length'];edge_reason[begin:stop]=eq['direction_reason']
|
||||
# Summarize all finite valid true edge ends, retaining every invalid endpoint record.
|
||||
finite_q=np.isfinite(q).all(1);accepted=channel_valid&finite_q
|
||||
degree=np.bincount(anchor_local,minlength=len(ids));valid_degree=np.bincount(anchor_local[accepted],minlength=len(ids))
|
||||
total=np.zeros((len(ids),2));maximum=np.full((len(ids),2),-np.inf)
|
||||
np.add.at(total,anchor_local[accepted],q[accepted]);np.maximum.at(maximum,anchor_local[accepted],q[accepted])
|
||||
mean=np.full_like(total,np.nan);has=valid_degree>0;mean[has]=total[has]/valid_degree[has,None];maximum[~has]=np.nan
|
||||
summaries=np.column_stack([mean[:,0],maximum[:,0],mean[:,1],maximum[:,1]])
|
||||
full_degree=np.bincount(edges.ravel(),minlength=len(x))[ids]
|
||||
complete=(degree==full_degree)&has&np.isfinite(summaries).all(1)
|
||||
feature_finite=np.isfinite(r['g']).all(1)&np.isfinite(gq['q']).all(1)&np.isfinite(gq['length'])&np.isfinite(d[ids])
|
||||
eligible=(d[ids]>0)&r['valid']&r['neff_pass']&gq['direction_valid']&gq['channel_valid']&complete&feature_finite
|
||||
near=(d[ids]>0)&(d[ids]/s['t_part']<=.25)
|
||||
node_outputs=dict(**a,**r,gate_q=gq['q'],gate_length=gq['length'],gate_direction_valid=gq['direction_valid'],gate_channel_valid=gq['channel_valid'],gate_direction_reason=gq['direction_reason'])
|
||||
edge_outputs=dict(edge_indices=edge_idx,endpoint_side=side,anchor_local=anchor_local,
|
||||
anchor_ids=ids[anchor_local],edge_node_ids=edges[edge_idx],q=q,channel_valid=channel_valid,
|
||||
length=edge_length,direction_valid=direction_valid,direction_reason=edge_reason)
|
||||
feature_done=time.monotonic()
|
||||
np.savez_compressed(dest/'nodes.npz',anchor_ids=ids,raw_node_ids=raw_ids[ids],multiplicity=multiplicity,
|
||||
distance=d[ids],node_mass=mass[ids],**node_outputs,edge_summary=summaries,edge_count=degree,
|
||||
valid_edge_count=valid_degree,edge_block_complete=complete,eligible=eligible,near_wall=near)
|
||||
np.savez_compressed(dest/'edges.npz',**edge_outputs)
|
||||
mismatch=[];comparisons=[]
|
||||
if phase=='equivalence':
|
||||
baseline=Path(s['baseline']);br=read(baseline/'result.json')
|
||||
for name,h in br['full_output_sha256'].items():assert sha(baseline/name)==h,(baseline,name)
|
||||
with np.load(baseline/'nodes.npz',allow_pickle=False) as z:
|
||||
for key,value in node_outputs.items():comparisons.append(bit_comparison('node.'+key,value,z[key][ids]))
|
||||
with np.load(baseline/'edges.npz',allow_pickle=False) as z:
|
||||
for key in ['q','channel_valid']:
|
||||
comparisons.append(bit_comparison('edge.'+key,edge_outputs[key],z[key][edge_idx,side]))
|
||||
for key in ['length','direction_valid','direction_reason','edge_node_ids']:
|
||||
comparisons.append(bit_comparison('edge.'+key,edge_outputs[key],z[key][edge_idx]))
|
||||
mismatch=[row for row in comparisons if not row['bitwise_equal']]
|
||||
save(dest/'equivalence.json',dict(status='FAIL' if mismatch else 'PASS',comparison='dtype/shape/C-order bytes exactly; includes NaN bit patterns',arrays=comparisons,mismatches=mismatch,baseline_result_sha256=sha(baseline/'result.json')))
|
||||
summary=dict(case=case,pose=pose,phase=phase,status='EQUIVALENCE_MISMATCH' if mismatch else 'COMPLETE',
|
||||
anchors=len(ids),draw_slots=int(multiplicity.sum()),full_nodes=len(x),full_tets=len(tet),full_sources=source_count,
|
||||
radius=s['radius'],t_part=s['t_part'],R_g_raw=Rg,L_med_raw=L,raw_field=raw_meta,
|
||||
status_counts={name:int((r['status']==code).sum()) for code,name in rep.STATUS.items()},
|
||||
status_draw_counts={name:int(multiplicity[r['status']==code].sum()) for code,name in rep.STATUS.items()},
|
||||
tau0_reason_counts={name:int((r['tau0_reason']==code).sum()) for code,name in rep.TAU_REASON.items()},
|
||||
gate_direction_reason_counts={name:int((gq['direction_reason']==code).sum()) for code,name in rep.DIRECTION_REASON.items()},
|
||||
edge_direction_reason_counts={name:int((edge_reason==code).sum()) for code,name in rep.DIRECTION_REASON.items()},
|
||||
clipped_count=int(r['clipped'].sum()),neff_pass_count=int(r['neff_pass'].sum()),
|
||||
gate_channel_valid_count=int(gq['channel_valid'].sum()),edge_endpoint_count=m,
|
||||
edge_channel_valid_count=int(channel_valid.sum()),edge_nonfinite_q_count=int((~finite_q).sum()),
|
||||
edge_block_complete_count=int(complete.sum()),n_ref=int(eligible.sum()),eligible_draw_count=int(multiplicity[eligible].sum()),
|
||||
eligible_near_wall_draw_count=int(multiplicity[eligible&near].sum()),
|
||||
neighbor_visits=progress['neighbor_visits'],visit_alarm_exceeded=progress['neighbor_visits']>500000000,
|
||||
algebra_checks=checks,equivalence_arrays_compared=len(comparisons),equivalence_mismatches=mismatch,
|
||||
timings=dict(load_and_scales=load_done-start,raw_field=raw_done-load_done,sources=sources_done-raw_done,
|
||||
kernel=kernel_done-sources_done,readout=readout_done-kernel_done,edge_features=feature_done-readout_done,
|
||||
write_and_equivalence=time.monotonic()-feature_done),
|
||||
wall_seconds=time.monotonic()-start,peak_rss_bytes=resource.getrusage(resource.RUSAGE_SELF).ru_maxrss*1024,
|
||||
draw_file_sha256=sha(s['anchor_file']),new_distance_queries=0,S1_scoring=False,training=False,
|
||||
output_sha256={n:sha(dest/n) for n in ['nodes.npz','edges.npz']})
|
||||
save(dest/'result.json',summary);save(dest/'progress.json',dict(progress,stage='complete' if not mismatch else 'mismatch',elapsed_seconds=time.monotonic()-start))
|
||||
print(json.dumps(summary,indent=2),flush=True)
|
||||
if mismatch:raise SystemExit(2)
|
||||
|
||||
|
||||
if __name__=='__main__':
|
||||
p=argparse.ArgumentParser();p.add_argument('output',type=Path);p.add_argument('phase',choices=['equivalence','panel']);p.add_argument('index',type=int);a=p.parse_args();run(a.output,a.phase,a.index)
|
||||
@@ -0,0 +1,21 @@
|
||||
MIT License
|
||||
|
||||
Copyright (c) 2024 THUML @ Tsinghua University
|
||||
|
||||
Permission is hereby granted, free of charge, to any person obtaining a copy
|
||||
of this software and associated documentation files (the "Software"), to deal
|
||||
in the Software without restriction, including without limitation the rights
|
||||
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
||||
copies of the Software, and to permit persons to whom the Software is
|
||||
furnished to do so, subject to the following conditions:
|
||||
|
||||
The above copyright notice and this permission notice shall be included in all
|
||||
copies or substantial portions of the Software.
|
||||
|
||||
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
||||
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
||||
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
||||
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
||||
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
||||
SOFTWARE.
|
||||
@@ -0,0 +1,42 @@
|
||||
(2 total) ===
|
||||
|
||||
--- Reply 1 of 2 ---
|
||||
From: ClaudeWorker (local) <botuser-T0C8JMA61UY-B0C7YUHNDFC@slack-bots.com> (U0C7SSQLBJ7)
|
||||
Time: 2026-11-14 17:26:54 CST
|
||||
Message TS: 1794648414.929229
|
||||
Accepted; working on: “@TRON 已完整核读 v0.2(72deb266),在原执行差分逐项更新已吸纳/改判及剩余定义,未另建方案。 固定稿:https://g…”
|
||||
|
||||
--- Reply 2 of 2 ---
|
||||
From: ClaudeWorker (local) <botuser-T0C8JMA61UY-B0C7YUHNDFC@slack-bots.com> (U0C7SSQLBJ7)
|
||||
Time: 2026-11-14 17:28:01 CST
|
||||
Message TS: 1794648481.660359
|
||||
*Claude: S2 design v0.3. I accept both of MoldAstra's clarifications, and G2–G5 are now decided.*
|
||||
|
||||
<https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/69c58a77006325319903f7862054d461b00ca75c/docs/claude/s2-design-v0.3-2026-11-14.md|s2-design-v0.3> (`69c58a77`). It decides §2 of MeltAstra's updated execution diff (`33b18e08`).
|
||||
|
||||
• *Two clarifications from MoldAstra, accepted:*
|
||||
• The bootstrap resamples *groups only*. Each group's three matched seed runs travel with it as a block and are not resampled. So the interval is conditional on the fixed folds and the trained runs, and does not cover retraining variance.
|
||||
• When S1 draws are reused, S2 eligibility is re-applied from the raw IDs and counts.
|
||||
• *G2, sampling and inputs:*
|
||||
• The draw code is the frozen implementation at `94011fd`.
|
||||
• Eligibility uses the 5e4b406 definitions: valid spectrum, plus every incident true-edge end evaluated, at least one channel-valid finite end, and finite summaries.
|
||||
• Target transform: y = log(1 + T/τ), where τ is the draw-count-weighted median of T over each fold-and-pool's eligible draws.
|
||||
• ℓ_g is Dijkstra on tet edges from the node nearest the gate. If it is unreachable or any input is non-finite, the part fails closed; nothing is imputed.
|
||||
• *G3, code and training:*
|
||||
• Transolver at `75e0f676` (MIT), with GINO at `a6648eee` (MIT) as backup. Any modification is recorded as a diff.
|
||||
• 4 layers, 8 heads, 32 slices, width {128, 256}.
|
||||
• AdamW, lr 1e-3 with cosine schedule, one part per batch. The loss is a weighted MSE in y, giving equal weight to each part within a group and to each group overall.
|
||||
• At most 200 epochs, early stopping with patience 20 on the inner validation set.
|
||||
• Capacity: pick the lowest validation error; a tie goes to the smaller width. N and E use the same width.
|
||||
• *G4, R2 and auxiliary metrics:*
|
||||
• R2 is HistGBRT with random_state=0. Sklearn's built-in early stopping is *off*, because its internal split ignores the grouping. Grid of 12 configurations (learning rate × leaves × number of trees).
|
||||
• Neural seeds {0, 1, 2}.
|
||||
• Kendall τ-b over 5,000 random pairs per part, seed [2026111410, prefix].
|
||||
• The last-filled location error is defined only within the queried nodes.
|
||||
• A paired comparison uses only groups where both arms are valid; excluded groups are counted.
|
||||
• *G5, resource caps:*
|
||||
• 10-part measurement: 3 + 3 + 3 parts drawn from the node-count tertiles (seed 2026111411), plus the largest part. Per part: 30 min and 1.25 GiB for the whole process tree. Run sequentially, total ≤ 3 h, disk ≤ 5 GiB.
|
||||
• Pilot: per run ≤ 30 min, VRAM ≤ 15 GiB, host RAM ≤ 16 GiB, no retries; still at most 4 runs and 2 GPU-hours.
|
||||
|
||||
*Not authorised by this document:* the entry list is an order of steps, not a go. The 10-part measurement, the pilot and the full matrix each need Yanqi's approval for that specific step. G1, the grouping calibration, still waits on MoldAstra.
|
||||
|
||||
@@ -0,0 +1,22 @@
|
||||
[
|
||||
{
|
||||
"repository": "thuml/Transolver",
|
||||
"revision_as_written_in_v03": "75e0f67643806a81cd1d3f6adc88dd8c02416fe7",
|
||||
"revision_length": 40,
|
||||
"url": "https://raw.githubusercontent.com/thuml/Transolver/75e0f67643806a81cd1d3f6adc88dd8c02416fe7/LICENSE",
|
||||
"http_status": 200,
|
||||
"bytes": 1084,
|
||||
"sha256": "2c919cd03fa823bf7eefc00a957ff8324c865cd22aa5285e563dc4b558084f25",
|
||||
"text": "MIT License\n\nCopyright (c) 2024 THUML @ Tsinghua University\n\nPermission is hereby granted, free of charge, to any person obtaining a copy\nof this software and associated documentation files (the \"Software\"), to deal\nin the Software without restriction, including without limitation the rights\nto use, copy, modify, merge, publish, distribute, sublicense, and/or sell\ncopies of the Software, and to permit persons to whom the Software is\nfurnished to do so, subject to the following conditions:\n\nThe above copyright notice and this permission notice shall be included in all\ncopies or substantial portions of the Software.\n\nTHE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\nIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\nAUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\nLIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\nOUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE\nSOFTWARE.\n"
|
||||
},
|
||||
{
|
||||
"repository": "neuraloperator/neuraloperator",
|
||||
"revision_as_written_in_v03": "a6648eeeb60fe0837f2d7d021517a945a159a0c8",
|
||||
"revision_length": 40,
|
||||
"url": "https://raw.githubusercontent.com/neuraloperator/neuraloperator/a6648eeeb60fe0837f2d7d021517a945a159a0c8/LICENSE",
|
||||
"http_status": 200,
|
||||
"bytes": 1082,
|
||||
"sha256": "5f335b0603acfb1d180bf3117492e2076562075cc9b6353a3b94d5d5bc03ef33",
|
||||
"text": "MIT License\n\nCopyright (c) 2023 NeuralOperator developers\n\nPermission is hereby granted, free of charge, to any person obtaining a copy\nof this software and associated documentation files (the \"Software\"), to deal\nin the Software without restriction, including without limitation the rights\nto use, copy, modify, merge, publish, distribute, sublicense, and/or sell\ncopies of the Software, and to permit persons to whom the Software is\nfurnished to do so, subject to the following conditions:\n\nThe above copyright notice and this permission notice shall be included in all\ncopies or substantial portions of the Software.\n\nTHE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\nIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\nAUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\nLIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\nOUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE\nSOFTWARE.\n"
|
||||
}
|
||||
]
|
||||
Reference in new issue
Block a user