Publish final R4bc three-pool results and independent bitwise determinism evidence
This commit is contained in:
1 parent
4bf84c59c8
commit
036290bb23
4166 files changed
+22851
No files matched your search
@@ -211,3 +211,6 @@ O1 v2最终审计收口:Claude1794865219接受53d95182及逐件分解。完整
|
||||
|
||||
|
||||
R4bc真实全拟合终态接续(2026-11-17):1005全部终态,1004PASS+原job90,105finalPASS,sourcepins/accountingPASS,无活动R4bc或现有postfit。按原合同进入检查点Git冻结→70个剩余模型预测(复用100%原预测)→完整预测/代码Git冻结→原评分→INTERIM_100零容差核对。原post14400与六小时共享时钟、原失败/费用完整继承;新独立编排7合成PASS/0.505717CPU秒、1004checkpoint哈希PASS。无重训、范围扩展或重复go;此条不声称预测/评分已运行,见docs/r4bc-final-postfit-2026-11-17.md及真实回执。
|
||||
|
||||
|
||||
R4bc最终全池完成(2026-11-17):105final全PASS,原job90 HPO失败保留。依Claude1794874951最终顺序,原70新预测加复用100%→Git冻结248542d→原全池score→INTERIM_100 EXACT_MATCH;随后独立35新预测3864数组BITWISE_MATCH。未重评分或改源码。三池五折齐全,所有主/链/相邻域552件204组件,未定义指标/bootstraps0。post1149.437161、总worker98516.406276CPU秒;全部进程终态。完整结果/原始预测与费用交Claude统一阶段收束,O1停止边界独立保留,不新增工艺/拓扑计算或用测试结果换模型。见docs/r4bc-final-results-2026-11-17.md。
|
||||
@@ -15,3 +15,6 @@ Latest controlling ruling: Claude1794874796.184549 supersedes1794874336.044949.
|
||||
|
||||
|
||||
Claude1794874951.316109 explicitly fixed the final sequence and withdrew the crossed earlier message. Executed accordingly: all-pool predictions locally frozen248542d23de8e737b50029b4502b577da4b098aa, original score guardPASS/38.284985CPU-s, exact INTERIM_100 comparePASS/7.881043CPU-s.35models/4968predictions plus100% part scores/summary/bootstrap EXACT_MATCH at zero tolerance. Main-chain post cumulative866.412064CPU-s. This alone did not test fresh100% prediction determinism because those entries were reused; the separately authorized35 check follows in artifacts/r4bc-determinism-100-2026-11-17. Its5 synthetic checksPASS; first fixture-only missing-directory failure0.271408CPU-s preserved, accepted attempt0.818123, both charged. No live or old source modified.
|
||||
|
||||
|
||||
FINAL SCIENTIFIC TERMINAL: the separate35 new predictions allPASS; bitwise3864 model arraysPASS, no defects. Main70+score+exact comparison alreadyPASS. All original/new coordinators absent at final host check. Post1149.437161/14400CPU-s and total worker98516.406276/144000; four metrics undefined0/bootstrap undefined0 across all552parts/204components and three pools. Scientific functions and old failures unchanged. Report docs/r4bc-final-results-2026-11-17.md. Original full training push54821 exited0 and remote05343e13 verified; final scores/determinism publication follows once, no duplicate training push. Do not restart/re-score/retrain.
|
||||
@@ -0,0 +1,109 @@
|
||||
# R4(b)/(c) 全三池终态与独立确定性核验
|
||||
|
||||
原1005拟合全部终态:1004PASS+原job90 HPO process_count_limit,105 final全部PASS。100%/50%/25%三池均完成五折外测,各项主表、全链及相邻边均552件/204组件;不是全原始数据域,也不是新的盲测。原失败及其1.269064CPU秒保留,无重训或重试。
|
||||
|
||||
**两个不同的一致性检查均通过。** 原70新预测+复用100%清单经Git冻结后,原scorer全池评分;35模型、4968条100%预测以及100%逐件分数/汇总/bootstrap与INTERIM_100零容差EXACT_MATCH。随后按Claude1794874951最终顺序,在独立目录重新调用原预测函数完成35个100%模型,3864份新模型预测的raw_ids/counts/seconds均在dtype、shape及数组字节上与interim一致,BITWISE_MATCH、defects=[]。独立35结果没有覆盖或重评分;其通过证明此前复用预测与此次新预测相同。没有把复用数据的自比较冒称预测确定性测试。
|
||||
|
||||
最后独立检查完成于2026-11-17T08:37:34.948017+08:00,在原04:06:06起算的六小时共享时限内。原模型、预测函数、评分与bootstrap函数、超参数及角色名单均未改。
|
||||
|
||||
## 主指标与学习池
|
||||
|
||||
下面是组件等权相对L2,数值越低越好。counts只计一次,组件内等件、组件间等权;与Gre24随机划分/全tet RRMSE不是同口径,不横向拼接数字。
|
||||
|
||||
|臂|100%池|50%池|25%池|
|
||||
|---|---:|---:|---:|
|
||||
|EIK|0.095879482537|0.095879482537|0.095879482537|
|
||||
|H0|0.073027071673|0.074655303044|0.076956158594|
|
||||
|H0p|0.073022990543|0.074978987490|0.077230813235|
|
||||
|H1|0.074262578885|0.075473231026|0.078264880648|
|
||||
|H2|0.072051346543|0.073182411712|0.076157691732|
|
||||
|H3|0.070024363499|0.071236875509|0.074341564189|
|
||||
|H4|0.070426230068|0.071815840211|0.074436876455|
|
||||
|A|0.070830540821|0.073081110105|0.077034763446|
|
||||
|Q|0.072781835831|0.074423417472|0.078093148168|
|
||||
|
||||
H0是历史参考;H0p即新H0′。H1删t_set输入,H2替换整套几何构造,H3在相同查询上改单位权,H4扩展到全节点单位权,A为固定XGB,Q为XGB查询层级权。H4→A同时包含学习器、float接口和HPO/固定参数政策差异;H0→H0′含名单和HPO响应,不作纯单因素效应。
|
||||
|
||||
## 配对差值与区间
|
||||
|
||||
Δ=前臂误差−后臂误差,正值表示后臂较低;是绝对差,不是相对提升百分比。95%区间来自原204组件配对bootstrap2000次、seed2026111408,条件于固定折/模型,作为描述性结果,不增设选择或通过门。
|
||||
|
||||
|池|比较|Δ|95%区间|
|
||||
|---|---|---:|---|
|
||||
|100%|EIK->A|0.025048941717|[0.022195926818, 0.027818408847]|
|
||||
|100%|EIK->Q|0.023097646706|[0.020387248607, 0.025891362837]|
|
||||
|100%|Q->A|0.001951295010|[0.001223714800, 0.002679049562]|
|
||||
|100%|H0->H0p|0.000004081129|[-0.000232486957, 0.000236333961]|
|
||||
|100%|H0p->H1|-0.001239588342|[-0.002015916781, -0.000492297742]|
|
||||
|100%|H1->H2|0.002211232342|[0.001466705021, 0.002872386974]|
|
||||
|100%|H2->H3|0.002026983044|[0.001414488614, 0.002665400914]|
|
||||
|100%|H3->H4|-0.000401866569|[-0.000940178294, 0.000165025748]|
|
||||
|100%|H4->A|-0.000404310752|[-0.000744109302, -0.000084007680]|
|
||||
|50%|EIK->A|0.022798372433|[0.019856719069, 0.025662240177]|
|
||||
|50%|EIK->Q|0.021456065066|[0.018838595010, 0.024181002810]|
|
||||
|50%|Q->A|0.001342307367|[0.000620763405, 0.002086472911]|
|
||||
|50%|H0->H0p|-0.000323684446|[-0.000565051341, -0.000096060611]|
|
||||
|50%|H0p->H1|-0.000494243536|[-0.001345177170, 0.000341746094]|
|
||||
|50%|H1->H2|0.002290819314|[0.001619497018, 0.002933663815]|
|
||||
|50%|H2->H3|0.001945536203|[0.001262305502, 0.002666009690]|
|
||||
|50%|H3->H4|-0.000578964702|[-0.001133460880, 0.000010292828]|
|
||||
|50%|H4->A|-0.001265269894|[-0.001701662066, -0.000856287750]|
|
||||
|25%|EIK->A|0.018844719091|[0.015819743742, 0.021692469498]|
|
||||
|25%|EIK->Q|0.017786334369|[0.014995325043, 0.020542703914]|
|
||||
|25%|Q->A|0.001058384722|[0.000319998803, 0.001760378455]|
|
||||
|25%|H0->H0p|-0.000274654641|[-0.000499979809, -0.000064949704]|
|
||||
|25%|H0p->H1|-0.001034067413|[-0.002205986901, -0.000017919649]|
|
||||
|25%|H1->H2|0.002107188916|[0.001344826397, 0.002843838055]|
|
||||
|25%|H2->H3|0.001816127542|[0.001089815111, 0.002588843565]|
|
||||
|25%|H3->H4|-0.000095312266|[-0.000665630491, 0.000471496417]|
|
||||
|25%|H4->A|-0.002597886991|[-0.003378172795, -0.001946827173]|
|
||||
|
||||
在这三个固定池中,A的主指标相对L2均低于Q及EIK,三个对应区间均为正。H1→H2和H2→H3也均为正;H3→H4的区间均跨零,末步H4→A均为负。删t_set的点估计在三个池均增加误差,50%池区间跨零,100%及25%池未跨零。H3点值在这张表里最低,但未据测试结果改模型选择、拟合、名单或后续实验。
|
||||
|
||||
学习池缩小时,A/Q及链中模型的点估计误差增加;三个池是嵌套学习集合,不当作独立重复。这里只描述固定公共域和已测试模型家族,不能推出普遍7%误差平台、其他工艺泛化、类型组织无效或真实连通路线失败。
|
||||
|
||||
## 辅助指标
|
||||
|
||||
τ-b越大越好;最后位置距离/Rg和近壁相对L2越低越好。下表直接读取同一冻结评分,不重算。
|
||||
|
||||
|池|臂|Kendall τ-b|最后位置/Rg|近壁相对L2|
|
||||
|---|---|---:|---:|---:|
|
||||
|100%|EIK|0.916525652665|0.342375611203|0.083638948330|
|
||||
|100%|H0|0.924479045724|0.441287552611|0.074931309083|
|
||||
|100%|H0p|0.924674448296|0.418977284803|0.074963075616|
|
||||
|100%|H1|0.923068425108|0.400495063672|0.076021959281|
|
||||
|100%|H2|0.925562040015|0.424134938005|0.074215496368|
|
||||
|100%|H3|0.927193927492|0.378983996603|0.071904839589|
|
||||
|100%|H4|0.927592410717|0.399624182857|0.072322301156|
|
||||
|100%|A|0.926649719537|0.376713807712|0.072787154312|
|
||||
|100%|Q|0.923894847343|0.434439880835|0.075026127258|
|
||||
|50%|EIK|0.916525652665|0.342375611203|0.083638948330|
|
||||
|50%|H0|0.923360888906|0.421230398424|0.076727548735|
|
||||
|50%|H0p|0.922924050512|0.429499436568|0.077080894126|
|
||||
|50%|H1|0.921897358781|0.402553872417|0.077579650707|
|
||||
|50%|H2|0.924488989608|0.423056988239|0.075649224765|
|
||||
|50%|H3|0.926012572242|0.379814468017|0.073301413533|
|
||||
|50%|H4|0.926272735033|0.417615187115|0.073806477133|
|
||||
|50%|A|0.924423917081|0.417964003936|0.075107000399|
|
||||
|50%|Q|0.922163137498|0.451455405911|0.076745720789|
|
||||
|25%|EIK|0.916525652665|0.342375611203|0.083638948330|
|
||||
|25%|H0|0.921206022397|0.433628907289|0.078845747896|
|
||||
|25%|H0p|0.920842255776|0.422387104995|0.078992355579|
|
||||
|25%|H1|0.919040223628|0.427568082992|0.080299053280|
|
||||
|25%|H2|0.921585181842|0.414549165871|0.078524087087|
|
||||
|25%|H3|0.922870856402|0.436800281680|0.076588888929|
|
||||
|25%|H4|0.924129021735|0.409006426280|0.076089365092|
|
||||
|25%|A|0.920483881366|0.420196300708|0.078679648878|
|
||||
|25%|Q|0.918484976902|0.453007301907|0.080092249615|
|
||||
|
||||
## 原始证据、费用与限制
|
||||
|
||||
主评分清单14904条:7728个新70模型预测、3864个原interim模型预测、3312条历史H0/EIK引用。最终评分14904行,四项指标未定义计数{'relative_l2': 0, 'kendall_tau_b': 0, 'last_location': 0, 'near_wall': 0};27个主/相邻比较的未定义bootstrap总数0。独立35检查另生成3864数组;与原70合计105个模型各完成一次本轮原函数预测,独立35没有进入第二次评分。全部108个新预测/评分/比较guard PASS,零GPU、零新拟合/VTU解码。
|
||||
|
||||
训练检查点/代码先固定05343e13e4ee386248461b6a1072899d42468d8c;主评分前完整预测固定248542d23de8e737b50029b4502b577da4b098aa;主评分与EXACT_MATCH及独立35计划固定4bf84c59c816585594490e2b56b2bff96e69d8d0。原科学pins不变。合成验收保留首个测试夹具目录错误,修复仅测试夹具,0.271408及0.818123CPU秒都在同post账;未伪装首次全通过。
|
||||
|
||||
原fit97267.128466CPU秒、setup99.840649CPU秒。全部post累计1149.437161/14400CPU秒(含原interim346.870387、新合成与90秒元数据保守预留),总worker累计98516.406276/144000CPU秒;不借O1或其他阶段预算,不重置共享时钟。原训练guard6096.497806331和协调10618.968657329CPU秒,以及各post协调费用单列。最末报告只读审计耗时0.678542CPU秒,纳入既有元数据预留。
|
||||
|
||||
机器结果:artifacts/r4bc-final-postfit-2026-11-17/scores/summary.json、scores/part-scores.json、INTERIM_100-exact-comparison.json、final-result-audit.json、final-result-tables.json。独立检查:artifacts/r4bc-determinism-100-2026-11-17/bitwise-comparison.json、complete.json及全部35新预测。所有原训练失败、HPO/模型、预测和访问日志保留共享。具体远端发布以回执为准,不把本地固定冒称上传完成。
|
||||
|
||||
研究目的仍是新件实用填充代理与样本效率;本表是固定公开域上的受控比较,不能替代新工艺或新浇口变化的验证。O1当前尺度因几何匹配覆盖不足而停止,与本R4bc终态分开解释。由Claude据共同证据写阶段收束和给Yanqi的建议,本次不自动新增实验。
|
||||
Reference in new issue
Block a user