Publish B17 terminal evidence and gate calibration v2 fixed execution scope
This commit is contained in:
1 parent
1e1d5a6515
commit
d67da6382b
57 files changed
+612570
No files matched your search
@@ -362,3 +362,21 @@ Claude1794913563明确仅要求pre-score身份PASS,Assist实际PASS SHA8cdd3ff
|
||||
|
||||
当前交Claude既有阶段回顾,无新实验授权。H2原587448/587587继续;B14/B15/E2终态不重跑。原第二bulk上传602200及transport-watch继续,B16正式包本地固定/Slack共享,远端发布待既有分流裁量,不另开同历史push。
|
||||
|
||||
|
||||
|
||||
## 最新 B17 终态与 gate outcome v2 接续(2026-11-17 20:20)
|
||||
|
||||
B17 已 COMPLETE / PRUNE_B17_FIXED_RECIPE,原121补充读出全部PASS,5fit/5predict全部PASS;552件/204组件、42 G-invalid保留。末点新−G均值−0.003285856649,95CI[−0.038931160490,+0.029789254516],预设均值≤−.010且CIhigh<0筛选未通过,stage2不启动;不否定独立类型/连通/其他loss路线。原657084于真实WAIT、无children、131终态/552预测hash核验后按Claude6398单次标记退役,原STOPPED和4677.737691664CPU-s保留。Assist PASS SHA235cfc9f在score前实读并固定;独立3da8a5a0入口launch28550589、664469/664500已退出0,0refit/0新预测。最终4744.511491480001/10800CPU-s,报告2f5dd7e8022a9c7155035f04db5d3df892607e15,结果SHA733af81182a0d7bfab6190ccbf9d6bd960f4886c6dc379d01f6d8cbb498d2f74。见docs/phase2-b17-results-2026-11-17.md;不重启/重评分。
|
||||
|
||||
已实读Claude1794916421:gate outcome v2独立72000CPU秒,52 exact gates hash复用、493固定原算法重算、7无效留分母;最多12worker避开B17CPU8/10/12/14,先10件成本投影,越72000停报无扩额。42候选下载继续暂停,exact-set FAIL保留。10合成验收PASS;准备元数据来源核对中,当前尚未启动真实v2特征计算,实际启动以后续真实started/audit为准。
|
||||
|
||||
已实读Claude1794917650:比较只用两gate预测均有定义的件;组件内对有定义件均值,组件等权;2000配对组件bootstrap、seed2026111408,不补0/插补。7/552报告资格损失;只有没有任何有定义件的组件从bootstrap排除并报告数目,并非要求整组件全件完整。两冻结模型field CI须在±.001、末点/Rg CI在±.010内。该规则在比较前固定,并须Assist事前语义PASS;现在不计算比较。
|
||||
|
||||
|
||||
|
||||
## 最新实际启动:gate outcome v2唯一准备队列(2026-11-17 20:22)
|
||||
|
||||
Claude6421/7650均实读;唯一service `meltmind-gate-outcome-v2-20261117`,coordinator666520,根`/home/frank14f/MeltMind-runs/phase2-gate-outcome-calibration-v2-2026-11-17`;launch3443ae534e22de0bfdfa602355abf7f9a69591af,合同5ea5b5bd06f9c40e57bf4f992a805ba42980223961ac1adbc94814e8f6473c35,comparison合同6b79ace13af01f29c4291231d880ca6c1fa4c0f063e009c4498eafcb8d8df69b。真实4guard+4worker、首批Di/Df/moments已见,首件已PASS;原固定前10成本门尚待。10合成PASS;52hash复用/493重算/7无效资格分母,12核上限,CPU16–27避开B17保留核,原单件1800秒/4GiB,总72000CPU秒含准备600保守记账和3600postfit储备。原先60秒准备预留在真实启动前更正,初版固定对象与无科学执行的元数据KeyError均保留。
|
||||
|
||||
队列仅特征准备,完整后状态FEATURES_COMPLETE_PREDICTIONS_AND_COMPARISON_PENDING;无fit/预测/比较入口,0候选标签读取,42下载仍暂停。比较规则已固定,后续实现仍须Assist实际语义PASS才比较,所有后续成本从同72000账继续。实际行为看started/live/first-ten/complete,不能从本快照重复启动或改pins。B17已终态PRUNE并在原Slack共享F0CJ09KV407,不重启。参见docs/phase2-gate-outcome-v2-execution-2026-11-17.md、actual-start-audit.json。
|
||||
|
||||
@@ -0,0 +1,13 @@
|
||||
# B17 stage 1: fixed last-fill screening
|
||||
|
||||
Claude1794915553.067999 accepts Mold b5bd8a4e as written. Root is `/home/frank14f/MeltMind-runs/phase2-b17-2026-11-17`, total10800 CPU-s, at most four workers, no GPU/retry/new v2 label access. The completed holdout calibration has priority and must precede this launch. H2-S1 continues unchanged. Stage2 is absent from this executable.
|
||||
|
||||
Reuse all431 existing B14 readouts byte-for-byte, including original missing/resource states; compute only the remaining121 from the original552/204 roster, using the unchanged B14 field function and 300 CPU/300 wall seconds per part. A defined resource timeout preserves its partial output and receives the original native-NaN missing representation; other engineering failures stop. Do not rerun the431. Before the remaining111 dispatch, the first10 new parts must pass a node-weighted1.25 cost projection. Geometry has6300 CPU-s, reserving4500 for postfit inside the same10800. Child CPU, guards and coordinator accounting are cumulative, with a120-second stop reserve. Root disk cap8GiB, per-worker RSS8GiB, aggregate32GiB; no retry or new budget extension.
|
||||
|
||||
Five original external folds, 100-percent learning pools, all552 test parts each once. Inputs23=original19+two frozen E1b G columns+two B14 K/S columns. Native-NaN and zero-family semantics are unchanged. Direct T/t_set target, XGB100/depth8/eta.1/seed0/one thread. Training-only last10-percent mass gets up to4:1 weight via exact tied-threshold allocation, normalized to mean-one node weight separately for each part. Constant-label parts receive unit weights. No weighted/label-dependent prediction inputs or post-hoc calibration.
|
||||
|
||||
Frozen G predictions are paired by query IDs/counts/coordinates/wall status and original fold/component. Score with original `part_metrics` and paired component bootstrap2000/seed2026111408. Screening passes only if mean candidate-minus-G last-location<=-.010 and95-percent upper endpoint<0. Otherwise prune this recipe (undefined outcomes remain PARK). Report field/tau and raw maximum tie counts; these cannot replace the primary gate. Passing means screened in, not confirmed.
|
||||
|
||||
After all predictions freeze, wait for actual MeltAssist review, frozen with the contract. The semantic pre-score gate requires exact PASS, the actual launch commit, contract SHA, launch-proof SHA and complete code SHA mapping. Missing verdict, FAIL, unrelated receipt or identity mismatch is rejected. No scorer may run on a generic file-freeze assertion alone. Assist must review the fixed implementation before scoring; source files remain unchanged while running.
|
||||
|
||||
New-work budgets do not charge historical B14 execution again. All431 reused inputs, existing failures, prior source hashes and the test roster are preserved. Any later confirmation requires qualified holdout features and at least25 new independent components; current min-T gate calibration failed and v2 eligibility is UNKNOWN. There is no automatic stage2 launch.
|
||||
@@ -0,0 +1,19 @@
|
||||
# B17 固定组合配方:筛选未通过(2026-11-17)
|
||||
|
||||
B17 stage1 COMPLETE / PRUNE_B17_FIXED_RECIPE。552件、204组件,全部原5折;42件G无效保持原生NaN输入,不删件。主指标为新配方减冻结E1b G的末点误差/Rg:均值需≤−0.010且95%组件配对区间上界<0。实际两个条件均未达到。此为固定23列+尾部加权配方的筛选,不是否定所有末点损失/路径/类型/连通路线,也不把训练组合解释为单因素因果效应。B17 stage2未获本结果放行。
|
||||
|
||||
|指标|B17|冻结G|B17−G|95%组件配对CI|
|
||||
|---|---:|---:|---:|---|
|
||||
|relative_l2|0.0700317036044|0.0697022143049|+0.000329489299548|[-1.77346985642e-05, +0.000676342337321]|
|
||||
|last_location|0.384303830928|0.387589687577|-0.00328585664917|[-0.0389311604904, +0.029789254516]|
|
||||
|kendall_tau_b|0.928088901233|0.928112243828|-2.33425942369e-05|[-0.00053152985221, +0.000552297693035]|
|
||||
|
||||
2000次组件bootstrap,seed2026111408,三个指标0个undefined replicate。逐件末点差描述性中位数为0;371个负查询预测保留,未裁剪。无验证集选型、无新标签解码。内部字段new_minus_A在此比较中的A代表冻结G,报告明确标为B17−G。
|
||||
|
||||
原121补充B14任务回执:{'PASS': 121};431旧读出按hash复用。5训练和5预测任务PASS,552预测先固定。原源码347073135ed79efaf5714374043bf46e13fc743d保持,已知描述性中位数误减dict缺陷保留。Claude1794916398批准独立修复3da8a5a0,仅取last_location.value、跳过None;12测试通过。Assist实读review PASS SHA235cfc9f29408be2de163a6ec399b679c3bc2e3957cc0cf71254200e97515149,发生在评分前;审阅仅实现/固定对象,不冒称独立数值复现。
|
||||
|
||||
原协调657084在真实WAIT、131终态且无子任务后仅一次带标签退役,原complete=STOPPED工程退役原样保留,非数值失败。原累计4677.737691664000 CPU-s计一次;原score与authorization未创建。独立评分根`/home/frank14f/MeltMind-runs/phase2-b17-reviewed-score-2026-11-17`,launch28550589d1728505011d8f40a9547046f3972e6b,合同fe3dc3b5988ebc559371475a2aa77f905ca9ec2da8ece3f3559231bb079555db。协调664469/worker664500均已退出,score guard PASS(6.35537 CPU-s、6.788018589 wall-s、183271424 B RSS),0重拟合、0重预测。
|
||||
|
||||
最终含terminal freeze累计4744.511491480001/10800 CPU-s。终态f8f11ccaecc9e72a881ba9936b300ea12462b736,结果SHA 733af81182a0d7bfab6190ccbf9d6bd960f4886c6dc379d01f6d8cbb498d2f74。terminal-audit.json重新核固定源、552预测hash、真实审阅时序、进程退出和成本,未重算统计。终态不重启/重评分。
|
||||
|
||||
进入决定:将结果交Claude现有阶段整合;独立gate outcome v2资格校准仍按既有72,000CPU秒裁量推进,不能由B17未通过推断该资格校准的结果。42候选下载继续暂停。
|
||||
@@ -0,0 +1,9 @@
|
||||
# B17 descriptive-median score repair
|
||||
|
||||
MeltAssist1794916379 correctly found that original34707313 subtracts two `last_location` metric dictionaries when constructing a descriptive per-part median. The real metric schema contains a nested `value`. This would raise TypeError after metric computation and before result output. No B17 scoring has occurred, and the original review authorization remains absent. This is an implementation defect, not a scientific result.
|
||||
|
||||
The original field/fit/predict implementation and its source pins remain unchanged. Independent `phase2_b17_score_repair.py` accesses the two `value` scalars, excludes undefined scalar pairs, and leaves an empty median null. The primary component summary, bootstrap, thresholds, model, predictions and query identity are unchanged. A synthetic test uses actual `part_metrics` objects and JSON result assembly; an AST comparison permits only the semantic review entry and descriptive-median input changes in the score body.
|
||||
|
||||
Claude1794916398.039119 approved this handoff, which follows the already established B16 engineering pattern: finish the121 field jobs and five fits/five prediction jobs, freeze all552 predictions, reach WAIT with no workers, then a labelled idle retirement of the original coordinator. Preserve its terminal receipt and charge its entire actual CPU once in the same10800 total. The new score-only entry cannot fit or predict and checks131 original records, all ten fitting/prediction PASS jobs, original channel/model/prediction freezes, source pins, and absence of the old score/authorization. No original failure or timeout is erased.
|
||||
|
||||
The new exact PASS receipt must bind the original34707313 commit, original contract and launch-proof SHA, plus the fixed repair commit and complete repair-source SHA map. Use the already fixed semantic validator from the B16 repair. No generic freeze-only permission and no post-guard gate is added. The handoff ruling is now recorded; actual PASS is still required before creating the score-only execution root or retiring the original idle coordinator. Preparation of this repair does not authorize scoring, and no active source is modified.
|
||||
@@ -0,0 +1,13 @@
|
||||
# Gate outcome calibration v2:固定执行与比较口径
|
||||
|
||||
Claude1794916421批准完整旧552件的独立72,000 CPU秒资格校准,1794917650明确比较统计。原exact-set校准FAIL不变;新的结果层面校准并非事后放宽旧门。42新候选下载暂停,候选标签不读取。本根在/home;不重新拟合模型,不改变原科学算法。
|
||||
|
||||
旧552分为52 exact seed set(原特征、G读出、预测按hash复用)、493非一致有效seed(固定tet-FIM Di/Df、梯度、moments、Vf与G读出重算)及7无效gate(保留资格分母)。Dw使用原始hash固定缓存;G使用原已固定分区,原G-invalid保持原生NaN,不再分区。候选gate只读取已封存raw IDs,不读取Fill_time数组。
|
||||
|
||||
唯一准备队列tools/phase2_gate_outcome_queue.py;先原固定顺序前10个重算件,用4worker测成本,node-weighted投影加准备开销与postfit储备,投影大于72000立即停止。通过后最多12worker、CPU16–27,保留B17的8/10/12/14;协调30–31。单件沿原1800 CPU秒/1800墙钟秒/4GiB树RSS/2GiB输出;全队列48GiB并发RSS、12GiB根目录。准备阶段最多68400CPU秒,保留3600给后续预测/比较;任意失败停止,无自动重试。
|
||||
|
||||
准备开销统一保守计600CPU秒,覆盖来源链查找、一次无科学计算的source_partition元数据KeyError、10项合成检查及最终固定。初版60秒预留在启动前更正,旧固定合同/proof原件保存在superseded-prelaunch,原总帽不增。旧holdout预算账68.367CPU秒保持独立,不从其余量补贴v2。10测试为3个与原算法位级一致的合成检查,7个计数/分母/CPU隔离/预算/成本边界/真实guard单次执行及失败保留检查,无候选数据。
|
||||
|
||||
比较合同comparison-contract.json在真实比较之前固定:两套冻结模型A、B13-E1b逐折绑定原checkpoint;仅使用label-gate与LOG-gate预测均有定义的成对件。组件内对有定义件取均值,再组件等权;2000次配对组件bootstrap、seed2026111408,不补0或插补。7/552明确报告资格损失;只有不存在任何有定义件的组件不进入bootstrap,报告排除组件数,不要求组件的每件都完整。两模型各自field CI均须在±0.001内、last-fill/Rg CI在±0.010内才通过。
|
||||
|
||||
当前准备入口只生成特征并停在FEATURES_COMPLETE_PREDICTIONS_AND_COMPARISON_PENDING,不实现拟合、预测或比较。后续预测及比较继续用同一72000账,全部预测先固定;真实Assist语义PASS必须在比较前读验/绑定。规则合同已固定不等于比较实现获审,也不等于校准通过。实际启动以根目录started/live/actual-start-audit为准。
|
||||
@@ -0,0 +1,9 @@
|
||||
# Holdout gate calibration: extractor not accepted for v2
|
||||
|
||||
The prespecified exact-node-set rule failed: **52/552 (9.420289855%)** agree with the existing LOG/YML gate mapping. All 552 parts were attempted; seven contain invalid full-field input under the frozen finite/nonnegative rule and emitted empty gate IDs. The remaining 493 nonmatching valid cases have a nonzero seed-node displacement. Maximum seed-set Hausdorff displacement is 0.9996374570280457 in original mesh-coordinate units. No label value, minimum, quantile or time field left the extractor; emitted per-part JSON contains gate_raw_ids only. No tolerance was selected after observing this result.
|
||||
|
||||
Both isolated extraction and the separate label-free comparison guards PASSed (1.69473 and 5.296685 CPU-s). The original gate mapping was independently reproduced from stored pre-solve gate coordinates and retained-node geometry for all 552. Whole-field Dw is the same cached tet-exterior field, whose boundary is independent of the gate; this is an algorithmic identity, not a new numerical solve. Differences in cached Dw at the proposed/reference seed sets were zero for all 545 valid extracted sets. Dw identity cannot rescue the failed exact-set acceptance rule.
|
||||
|
||||
Frozen implementation/contract: 5fadea9e72154fcebcaf80f782210fc420a2f648 / 56c930d078610bd50def931f57b1f3135b6e4ace507112b8799f8f3b198a8027. Terminal: 6298448eba45458fc0e2553a29261dcecd364a31. Coordinator654052 and workers654056/654065 completed. Cumulative holdout cost including identity preparation and 60-second setup reserve is 68.367046134 / 7200 CPU-s; there are no new v2 label reads, fits, predictions or scores.
|
||||
|
||||
Per Claude1794915157, v2 feature eligibility remains UNKNOWN and the extractor is not run on v2. The 94/42/40 candidate identity seal remains intact. Geometry and near-duplicate eligibility remain UNKNOWN; the 42-member public fetch is authorized but has not started because the gate-calibration completion criterion now exposes the concrete input blocker. This is not a claim that the geometries or simulations are invalid. A qualified fresh holdout has not been completed, and B17 stage2 has no usable confirmation cohort. No guessing, threshold relaxation or unsealing follows from this result.
|
||||
@@ -1,3 +1,17 @@
|
||||
## 最新实际启动:gate outcome v2唯一准备队列(2026-11-17 20:22)
|
||||
|
||||
Claude6421/7650均实读;唯一service `meltmind-gate-outcome-v2-20261117`,coordinator666520,根`/home/frank14f/MeltMind-runs/phase2-gate-outcome-calibration-v2-2026-11-17`;launch3443ae534e22de0bfdfa602355abf7f9a69591af,合同5ea5b5bd06f9c40e57bf4f992a805ba42980223961ac1adbc94814e8f6473c35,comparison合同6b79ace13af01f29c4291231d880ca6c1fa4c0f063e009c4498eafcb8d8df69b。真实4guard+4worker、首批Di/Df/moments已见,首件已PASS;原固定前10成本门尚待。10合成PASS;52hash复用/493重算/7无效资格分母,12核上限,CPU16–27避开B17保留核,原单件1800秒/4GiB,总72000CPU秒含准备600保守记账和3600postfit储备。原先60秒准备预留在真实启动前更正,初版固定对象与无科学执行的元数据KeyError均保留。
|
||||
|
||||
队列仅特征准备,完整后状态FEATURES_COMPLETE_PREDICTIONS_AND_COMPARISON_PENDING;无fit/预测/比较入口,0候选标签读取,42下载仍暂停。比较规则已固定,后续实现仍须Assist实际语义PASS才比较,所有后续成本从同72000账继续。实际行为看started/live/first-ten/complete,不能从本快照重复启动或改pins。B17已终态PRUNE并在原Slack共享F0CJ09KV407,不重启。参见docs/phase2-gate-outcome-v2-execution-2026-11-17.md、actual-start-audit.json。
|
||||
|
||||
## 最新 B17 终态与 gate outcome v2 接续(2026-11-17 20:20)
|
||||
|
||||
B17 已 COMPLETE / PRUNE_B17_FIXED_RECIPE,原121补充读出全部PASS,5fit/5predict全部PASS;552件/204组件、42 G-invalid保留。末点新−G均值−0.003285856649,95CI[−0.038931160490,+0.029789254516],预设均值≤−.010且CIhigh<0筛选未通过,stage2不启动;不否定独立类型/连通/其他loss路线。原657084于真实WAIT、无children、131终态/552预测hash核验后按Claude6398单次标记退役,原STOPPED和4677.737691664CPU-s保留。Assist PASS SHA235cfc9f在score前实读并固定;独立3da8a5a0入口launch28550589、664469/664500已退出0,0refit/0新预测。最终4744.511491480001/10800CPU-s,报告2f5dd7e8022a9c7155035f04db5d3df892607e15,结果SHA733af81182a0d7bfab6190ccbf9d6bd960f4886c6dc379d01f6d8cbb498d2f74。见docs/phase2-b17-results-2026-11-17.md;不重启/重评分。
|
||||
|
||||
已实读Claude1794916421:gate outcome v2独立72000CPU秒,52 exact gates hash复用、493固定原算法重算、7无效留分母;最多12worker避开B17CPU8/10/12/14,先10件成本投影,越72000停报无扩额。42候选下载继续暂停,exact-set FAIL保留。10合成验收PASS;准备元数据来源核对中,当前尚未启动真实v2特征计算,实际启动以后续真实started/audit为准。
|
||||
|
||||
已实读Claude1794917650:比较只用两gate预测均有定义的件;组件内对有定义件均值,组件等权;2000配对组件bootstrap、seed2026111408,不补0/插补。7/552报告资格损失;只有没有任何有定义件的组件从bootstrap排除并报告数目,并非要求整组件全件完整。两冻结模型field CI须在±.001、末点/Rg CI在±.010内。该规则在比较前固定,并须Assist事前语义PASS;现在不计算比较。
|
||||
|
||||
## 最新 B16 COMPLETE:固定组合配方 PRUNE(2026-11-17 19:14)
|
||||
|
||||
Claude1794913563明确仅要求pre-score身份PASS,Assist实际PASS SHA8cdd3ff42f57ea1b6005e8ce47815edcc59dbc856af502328dfc7463bc5f2eb1先于评分读取/固定。72d45697四源不改;v2草稿弃用、不进authorization、不运行。原前后门措辞以独立docs/phase2-b16-score-gate-wording-erratum-2026-11-17.md更正,终态hash仅事后审计。NEEDS原件保留。
|
||||
|
||||
Reference in new issue
Block a user