Preserve terminal seed1 continuation per-fold projection stop and full raw evidence

This commit is contained in:
MeltAstra committed 2026-11-18 16:34:19 +08:00
1 parent 90085f1842
commit d3c94a0069
122 files changed
+170997 -45

No files matched your search

@@ -655,3 +655,8 @@ Claude1794984996及1794985731已实读:一次独立root接续,同seed1/200ep
Seed1单次接续实际启动:Assist新PASS5c8359c9按75eff37b/e90bc1b5完整绑定已采用;2026-11-18 15:16唯一coord1389930/fold0worker1390020在GPU0,实际三idle polls和worker进度核验。GPU1外来占用被正确绕开,Track A原队列保持。旧539.072057内扣新余额,累计167200和phase3612000不变;首门待实测,未经结果不启动后续阶段。详见continuation-launch报告和真实receipts。
## Seed1接续终态与后续新设计边界(2026-11-18 16:26:59)
原单次接续在fold1 epoch34投影31840.685947>余帽31730.464056停止,总投影158673.914420仍低于166660.927943;两代seed1进程已全退出,new实际7943.241351+old539.072057一次。源及raw证据冻结e3f25148/audit3bdaf94c;无科学性能结论或慢轮原因归因。Track A继续。Claude1794990750另提出相同累计167200内v2,变更总投影/40000 runaway和epoch外来进程观测,要求独立新绑定窄审;不等于复活本次终态。终态报告与新设计分开,旧STOP/失败永存。当前核验没有v2实现/训练或后处理。
+4
View File
@@ -1,3 +1,7 @@
## 最新 seed1接续已终态STOP,禁止按首门快照续称运行(2026-11-18 16:26:59)
唯一1389930及所有两代seed1进程退出,service16:27:03 exit1。fold1 epoch34训练134.211354s导致投影31840.685947>单折新余帽31730.464056;新总投影158673.914420<166660.927943,不能称总预算耗尽。fold0完成46轮、fold1完成34轮,下一轮中途保护杀停;2/3/4未尝试,无checkpoint/postfit。new实际7943.241351+old539.072057一次=累计seed18482.313408,phase3实际80903.629407。guard monitor[]/worker int0、return−9、无内存越界/数值failure;慢轮外部原因未建立。75eff37b/e90bc1b5/5c8359c9及原STOP源保持。budget-stop audit3bdaf94c、freezee3f25148;报告docs/phase3-seed1-continuation-stop-2026-11-18.md。Track A原1262364继续,前两PASS、2/3核时epoch48,未干扰。Claude1794990750另提同总帽v2需新绑定及独立逻辑审阅;本次终态核验没有恢复、扩帽、v2实现或启动。下方所有运行/首门快照为历史。
## 最新 接续fold0首门PASS,原队列继续(2026-11-18 15:17:56)
同一1389930/1390020,首轮94.436642s,fold0投影24091.671488低于新余额31730.463887。新run总投影159022.135544含其余四折整帽+7200postfit;加旧539.072057一次,累计seed1投影159561.207601/167200,phase3投影231982.523600/612000。实际已进入epoch2;pending四折尚未实测,不称五折成本全PASS。first-dispatched-epoch-audit及固定源码保持;不新增launch/重试/评分,之后继续逐epoch保护,终态以raw receipts为准。
@@ -0,0 +1,31 @@
# Seed1 continuation: terminal per-fold projection stop
The sole continuation stopped at 2026-11-18 16:26:59 Asia/Shanghai (Unix1794990419.773086), reason PROJECTED_SEED1_BUDGET. Service exited1 at16:27:03; independent host inspection found no original or continuation seed1 queue/guard/worker. Track A's original coordinator1262364 continues, first two cells PASS and cells2/3 active at epoch48 in this audit. Track A was not changed.
The binding remains implementation75eff37b9a2ba656adc3617e11a56f51384ff666 / contracte90bc1b5e3781b22d2aed4b7a9b4f1f6b64ca0eebdcb1c6fa7eb42696d9752b0 / review5c8359c9. Source hashes, exact review binding and Track A pins passed verification. The original seed1 STOP and its539.0720570000121 GPU-s remain immutable.
## Trigger and interpretation
Fold1 completed epoch34 with train_epoch_seconds134.2113543249434, its new maximum. Under the unchanged max-epoch ×1.25 formula, its projection31840.68594737444 exceeded remaining per-fold31730.46405574202 by110.22189163242 GPU-s. Fold0 projection23633.22847286641 was within its cap. The new-run total projection158673.91442024085 was **below** its166660.927943 cap. Adding old539.072057 once gives cumulative seed1 projection159212.98647724086 below167200. Phase3 projection231634.3024766277 was below612000. Thus the per-fold projection condition triggered; neither total projected allowance nor actual total budget was exhausted.
This is an incomplete run stopped by the agreed resource projection, not a numerical/scientific-performance failure. The max-duration increase is observed; its underlying cause is not established by retained samples. A current or startup foreign-process snapshot does not prove interference during epoch34. Claude1794990750's shared-host attribution remains an interpretation rather than an independently established cause. No new metric analysis or recomputation was performed.
## Work and actual charges
| Fold | Completed epochs | Partial next epoch | Optimizer steps | New GPU-s charged |
|---|---:|---:|---:|---:|
|0|46|47,240/320 parts|14960|4317.711953648948|
|1|34|35,2/374 parts|12718|3625.5293972480576|
|2/3/4|0|Never attempted|0|0|
New-run actual7943.241350897006 GPU-s + original539.0720570000121 exactly once = cumulative seed18482.313407897018 GPU-s. Original item1 actual72421.31599938683 plus cumulative seed1 = phase3 actual80903.62940728385. These actual charges are distinct from all projections above. Postfit spent0, original7200 reserve unused.
New worker CPU sum7800.270604000001s; enclosing guard CPU charges8735.409249683002s already include worker CPU and setup allowances, so do not add them together. The stopped coordinator does not serialize its CPU total; this limitation is retained, without invented estimates.
Both guard reasons are exactly SEED1_GLOBAL_STOP, COORDINATOR_SIGNAL_15, WORKER_NOT_PASS, NO_COMPLETE_TRAIN_RESULT. Wrapper monitor_errors=[] and worker monitor_errors=integer0. Worker return−9 is consistent with the reviewed guard SIGKILL stop path; wrapper return1. Recorded loss/gradient values were finite, no numerical failure receipt, no memory breach. Peak worker RSS1458262016/1468104704B; reserved VRAM1388314624/1390411776B. Sampled maximum poll gaps2.583525/2.183840s against target0.2s are retained; not represented as continuous observation. No checkpoint, final model result, predictions or scoring exists. Original partial histories and raw monitoring evidence remain intact.
## Preservation and next-stage boundary
Root `/home/frank14f/MeltMind-runs/phase3-seed1-continuation-2026-11-18/budget-stop-audit.json` SHA3bdaf94c4b4c211fc46ae156b08136aac07736e4bf13d42d8c4e9d20ce0d919f; immutable local budget-stop freezee3f251489d273037f47300ac295c174237551b7b. Audit rechecks guard/assignment/terminal hashes, contiguous histories, last partial progress, original-charge accounting, absence of later-fold directories, process exit and source identity. Projection reconstructed from retained epoch/validation maxima and assignment wall time matches the recorded result within0.1s; this is metadata verification, not training or metrics rerun.
This run is terminal and is not extended or resumed. Claude1794990750 has separately proposed seed1 v2 under the same167200 cumulative total, debiting both prior attempts once, changing the stopping logic to aggregate projection plus40000/fold runaway ceiling and recording epoch-level foreign GPU observations. That is a distinct future binding requiring its stated independent review, not an action performed in this terminal audit. No v2 launch, parameter change, source mutation or postfit occurred here. The present report and all previous failed attempts remain regardless of later decisions.