Record fixed seed1 v2 actual launch with exact independent PASS
This commit is contained in:
1 parent
2dc2f39e37
commit
38f0044299
68 files changed
+11324
No files matched your search
@@ -668,3 +668,8 @@ Claude1794990899/0914明确:pending预留started当前投影median,无实测
|
||||
|
||||
|
||||
Seed1 v2同次独立审阅发现观测覆盖P2(Assist1794991467,BLOCKEDfe873785),非预算或科学失败。原绑定/56文件/回执不改而保存快照。最小三文件修复已固定8603801c/c8077c38,18CPU检查含终态早于poll、双receipt积压、global-stop与worker消失;仅修观察coverage,已通过的账本/停止公式保持。新实际review仍须精确五字段绑定;当前待同次复核,不重复主审、不启动训练或干扰Track A。
|
||||
|
||||
|
||||
## V2版本选择与实际启动(2026-11-18 16:55:39)
|
||||
|
||||
旧版waiver1794991491在执行前由Claude1794991926明确改为修复版8603801c+同次delta实际PASS,不将waiver跨绑定。Assist新PASSa585c24c已采用并验证完整绑定;唯一1476593/worker1476707和1476726实际GPU0/1 epoch1,初始成本门待完成。原BLOCKED/两版快照/两STOP保持,未建立waiver入口。旧8482.313408一次、总167200不变;Track A独立继续。详见v2-launch报告及实际receipts,不以启动代结果或提前后处理。
|
||||
@@ -1,3 +1,11 @@
|
||||
## 最新 v2首两成本门和观测回执PASS(2026-11-18 16:57:43)
|
||||
|
||||
同一1476593队列/1476707、1476726 worker,首轮88.252592/106.888871s,成本门均PASS且已epoch2;两实际观测OBSERVED/errors[]/history hash相符,不倒推历史占用。pending2/3/4各按median24863.510897预留,新总投影131517.554487含post7200,加旧8482.313408一次=seed1累计投影139999.867895/167200,phase3投影212421.183895/612000。三pending尚未实测,不称全五折PASS。source8603801c/c8077c38不变,first-two-epoch-audit记录原回执;不复制、不改帽、不自动后处理。
|
||||
|
||||
## 最新 修复版seed1 v2已真实启动(2026-11-18 16:55:39)
|
||||
|
||||
Claude1794991926选8603801c,废弃旧版waiver路径;Assist新PASSa585c24c/evidencec417a72c原样采用freeze/validate,原BLOCKEDfe873785不改。唯一service meltmind-phase3-seed1-v2-20261118(Restart=no),coord1476593;fold0 guard1476627/worker1476707 GPU0,fold1 guard1476628/worker1476726 GPU1,真实epoch1。三idle polls/assignment匹配,初次成本门尚待完成。绑定8603801c/c8077c38/baa0f87e/cba51291/51map及407源保持。old8482.313408在167200内一次,新cap158717.686592/train151517.686592+post7200,phase3prior80903.629407;actual40000 runaway/总投影median规则不变、无进一步seed1接续。RAM52.215GB/disk237.658GB通过,Track A原1262364及2/3卡不变、其source pins PASS。首个只读audit早于progress写出已保留attempt,之后实见两worker再记录actual-start,无重启。报告docs/phase3-seed1-v2-launch-2026-11-18.md;当前状态看raw live/terminal,不复制队列/改固定源/自动后处理。下方待复核、waiver及原两STOP均历史。
|
||||
|
||||
## 最新 v2观测覆盖阻断已修复固定,待同次复核(2026-11-18)
|
||||
|
||||
Assist原BLOCKEDfe873785仅V2-EPOCH-OBSERVATION-COVERAGE:terminal先collect移除及rows[-1]导致漏最后/积压epoch;原预算/ledger/科学通过。原56文件/绑定/复现及BLOCKED完整保存在artifacts/phase3-seed1-v2-observation-fix-2026-11-18。只改observation/queue/test三文件,新implementation8603801c/contractc8077c38/proofbaa0f87e/comparisoncba51291/51codes、407sources PASS,18CPU检查PASS。独立coverage游标、终态/停止清理前逐条排空;迟到/积压/terminal标INCOMPLETE不伪造历史,当前snapshot真实起止,worker消失显式INCOMPLETE。common预算/投影/idle及worker/guard不改。新binding已请求TRON转Assist同次复核,原BLOCKED不得启动;无review采用/started/launch。报告docs/phase3-seed1-v2-observation-fix-2026-11-18.md。两旧STOP与Track A不变,下方原送审c76719ff为历史。
|
||||
|
||||
@@ -0,0 +1,18 @@
|
||||
# Fixed seed1 v2 actual launch
|
||||
|
||||
Claude1794991926.661579 selected the fixed implementation8603801c and superseded the old-version waiver path. Actual MeltAssist delta PASS a585c24ceb53fe9fcf8d6b919fe1ac5a42f262c1cfdcf86ddd487a8fd4e76349 (evidencec417a72cfbb1133a9fe8721d8dc75dbc78b1685d7f4fe3187070b9557586a42c) closes V2-EPOCH-OBSERVATION-COVERAGE. The PASS was adopted byte-for-byte, frozen and validate_review passed. Original BLOCKEDfe873785, original c76719ff snapshot and both prior STOPs remain. No waiver adapter was created or used; original code was not restored.
|
||||
|
||||
The sole service `meltmind-phase3-seed1-v2-20261118.service` started at2026-11-18 16:55:39 Asia/Shanghai, Restart=no. Coordinator1476593 dispatched fold0 guard1476627/worker1476707 toGPU0, and fold1 guard1476628/worker1476726 toGPU1. Independent actual-start inspection observed both workers with real epoch1 progress; both assignment SHA values match guard/worker commands and three idle polls passed. This confirms launch, not completion or scientific performance. First-epoch cost gates were not yet complete at this initial audit.
|
||||
|
||||
Root `/home/frank14f/MeltMind-runs/phase3-seed1-v2-2026-11-18`: implementation8603801c4644068db5102227878e1af28c3b5f5e, contractc8077c380802615f1caed9ca28448ad8b770e5dd83dc7edb4c0a618a3ad15981, proofbaa0f87e82ff1cb36e6c09ee7e4d0fabe96808a3833776196bd21e53b7571d17, comparisoncba51291811c224966ccab82ff64a9b0690c83eaf2482316853ec9c656322493;51 code/407 source pins unchanged. Common budget/project/dispatch logic remains as reviewed, with independent observation coverage and explicit incomplete records described in the repair report.
|
||||
|
||||
Startup host audit: available RAM52215209984B >36507222016B required for the initial two workers plus2GiB; home free237657681920B >100GiB. GPU0/1 each5MiB and0% utilization. GPU2/3 remained with Track A workers1416483/1414160 under original coordinator1262364; Track A source pins independently passed. Load average15.3228/7.0933/5.3618 and CPU45.3% on32 logical CPUs were recorded without attributing historical slow epochs. No foreign workload was modified, and GPU2 remains reserved for Track A postfit.
|
||||
|
||||
Old539.0720570000121+7943.241350897006=8482.313407897018 is charged once within cumulative167200. New v2 total158717.68659210298 includes training151517.68659210298 and7200 postfit. phase3prior80903.62940728385 already includes both prior seed1 costs; only new v2 cost is added. Initial actual-start snapshot new-run charge201.9295999010792 gives seed1 cumulative8684.243007798097 and phase3 cumulative81105.55900718493; these are time-stamped actuals, not forecasts or final charges. Actual per-fold runaway40000, total max-epoch×1.25 projection with pending median, and no further continuation remain fixed.
|
||||
|
||||
The first read-only start audit ran before fold0 progress.json had been written and did not produce an actual-start claim. `actual-start-audit-attempt-01.json` retains that missing-progress observation; after both receipts appeared the read-only audit succeeded. No training launch was repeated.
|
||||
|
||||
Evidence includes root review-adoption, implementation-review freeze, host-precheck, launch-attempt/service-launch, started, actual-start-audit and per-fold assignment/idle/progress. Stage remains training only; no checkpoint result, prediction, scoring or downstream item1b is claimed. Current state must be read from actual live/terminal receipts. A completed epoch with late/terminal/backlog observation remains explicitly incomplete; current snapshots are never historical back-fill.
|
||||
|
||||
|
||||
First-two milestone (16:57:43): fold0/1 first-epoch train88.252592/106.888871s, both cost gatesPASS and actual epoch2 observed. Their projections22547.671018/27179.350777 give pending2/3/4 median reserve24863.510897 each; including7200postfit, new-run total projection131517.554487, plus old8482.313408 once = cumulative seed1139999.867895 below167200. Phase3 projection212421.183895 below612000. Each first-epoch observation is actually OBSERVED, errors[], with true snapshot start/end and matching history SHA; these are current snapshots, not continuous or historical occupancy proof. First-two-epoch-audit retains all receipts. Three pending folds remain unmeasured; no five-fold cost/performance PASS is claimed.
|
||||
Reference in new issue
Block a user