Inventory existing G1 material and validate virtual scheduler contract
This commit is contained in:
1 parent
07003020c5
commit
24d2da0fe7
14 files changed
+662
No files matched your search
@@ -0,0 +1,43 @@
|
||||
{
|
||||
"status": "INVENTORY_IDENTITY_FILES_SEALED_LOCALLY_NOT_BLIND_PACKAGE",
|
||||
"custodian": "MeltAstra",
|
||||
"reviewer": "MoldAstra",
|
||||
"unblinding_witness": "Grok",
|
||||
"source": "https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794654335163209",
|
||||
"identity_directory": "artifacts/s2-g1-inventory-2026-11-14/custodian",
|
||||
"directory_mode": "0700",
|
||||
"file_modes": "0600",
|
||||
"files": [
|
||||
{
|
||||
"file": "independent_natural_pairs.json",
|
||||
"bytes": 117,
|
||||
"sha256": "5ccb2981e7be4669f9c9d28e719e5d3ac0c2c123e15094ea1c2d5b4b4260e6ba"
|
||||
},
|
||||
{
|
||||
"file": "manifest.json",
|
||||
"bytes": 588,
|
||||
"sha256": "7a0bf4e91180a6745d18e0570510cec0744818ca5f2dd47a03e71388b96708d9"
|
||||
},
|
||||
{
|
||||
"file": "old_exposed_pairs.json",
|
||||
"bytes": 13164,
|
||||
"sha256": "cb66a776c89887fa76fa9a43dc14801a7a8c43a94cb37b2c7b72a3da84dc5955"
|
||||
},
|
||||
{
|
||||
"file": "parts.json",
|
||||
"bytes": 900882,
|
||||
"sha256": "977df4e1910607a2b540dce4a4d95dff247414c88daa54415a6799feb71abb35"
|
||||
},
|
||||
{
|
||||
"file": "source_clusters.json",
|
||||
"bytes": 86487,
|
||||
"sha256": "b2fd8451bfefede6d11523c118f620c9748964cb9c17b06a2c6bfa8f3ec97a2d"
|
||||
}
|
||||
],
|
||||
"no_anonymous_mapping_created": true,
|
||||
"no_sampling_or_probability_assignment": true,
|
||||
"no_blind_package_dispatched": true,
|
||||
"supersedes_inventory_summary_custody_pending_status": true,
|
||||
"unblinding": "freeze and commit blind labels first; Grok verifies published SHA256 before joining identities. No seal opened in this inventory stage.",
|
||||
"limitations": "Unix permissions are local access control, not encrypted storage or protection from owner/root; no claim of technical isolation from authorized shared-host agents."
|
||||
}
|
||||
@@ -0,0 +1,9 @@
|
||||
{
|
||||
"parts_unique": 624,
|
||||
"clusters_cover_all_parts": true,
|
||||
"old_pairs_retained": 30,
|
||||
"no_fabricated_parent_or_positive_truth": true,
|
||||
"metadata_export_allowlist_checked": true,
|
||||
"surface_count_reconciled": true,
|
||||
"no_blind_view_generated": true
|
||||
}
|
||||
@@ -0,0 +1,13 @@
|
||||
{
|
||||
"baseline_commit": "108e78364437c5f79304610df6c63dd4daad5e76",
|
||||
"baseline_e1_charged_cpu_seconds": 1665.366982,
|
||||
"new_preparation_reserve_seconds": 30,
|
||||
"new_reserve_is_measurement": false,
|
||||
"new_reserve_scope": "metadata inventory, synthetic model tests including initial failure, seals and publication; measured receipts included within reserve, not added again",
|
||||
"measured_inventory_cpu_seconds": 0.310497,
|
||||
"measured_final_test_process_cpu_seconds": 0.08585,
|
||||
"updated_conservative_e1_cpu_seconds": 1695.366982,
|
||||
"e1_limit_cpu_seconds": 57600,
|
||||
"scientific_workers_started": 0,
|
||||
"registration_queue_started": false
|
||||
}
|
||||
@@ -0,0 +1,70 @@
|
||||
{
|
||||
"scope": "public aggregates/code/simulation receipts and seal hashes; actual identity files remain on Horsea under custodian",
|
||||
"files": [
|
||||
{
|
||||
"path": "artifacts/s2-g1-inventory-2026-11-14/public/custody-seal.json",
|
||||
"bytes": 1693,
|
||||
"sha256": "bb7695b532a1db1cb2180db93d64408807b7c7db37462780c515bc48da069ee5"
|
||||
},
|
||||
{
|
||||
"path": "artifacts/s2-g1-inventory-2026-11-14/public/inventory-validation.json",
|
||||
"bytes": 256,
|
||||
"sha256": "c715ce3d2850e5405906a46f08a60c54cdd804d55f32fe23731c26693189f36c"
|
||||
},
|
||||
{
|
||||
"path": "artifacts/s2-g1-inventory-2026-11-14/public/preparation-resource-ledger.json",
|
||||
"bytes": 636,
|
||||
"sha256": "9f2c29f6e10e4b542436a626c468c9f9ba717fa5ced0bdcf852944f296df3f1f"
|
||||
},
|
||||
{
|
||||
"path": "artifacts/s2-g1-inventory-2026-11-14/public/summary.json",
|
||||
"bytes": 2134,
|
||||
"sha256": "7a8e9fe4c4b7b4007cf278093322ad56ac340bb6857082bdcb51070b6ef6fa5d"
|
||||
},
|
||||
{
|
||||
"path": "artifacts/s2-g1-scheduler-simulation-2026-11-14/initial-simulator.py",
|
||||
"bytes": 5929,
|
||||
"sha256": "5b09c8e7288684cc07a2b789e0905ef3905a9030ea5778931a1b52212d8c4f21"
|
||||
},
|
||||
{
|
||||
"path": "artifacts/s2-g1-scheduler-simulation-2026-11-14/initial-test-failure.json",
|
||||
"bytes": 387,
|
||||
"sha256": "71dd9a96bb17b93455ae7066109738bd520c03f8187195bc8fd9c3714f96bad2"
|
||||
},
|
||||
{
|
||||
"path": "artifacts/s2-g1-scheduler-simulation-2026-11-14/test-receipt.json",
|
||||
"bytes": 486,
|
||||
"sha256": "ebf9900623cb497cf52e75972a690e53947e08ab477f28ba097a24cd4aa9ea10"
|
||||
},
|
||||
{
|
||||
"path": "artifacts/s2-g1-scheduler-simulation-2026-11-14/tests.log",
|
||||
"bytes": 1063,
|
||||
"sha256": "c5283afa5a60b8e7b7296f838feee8cc5c837b1c28187e7097d8eaaeda846af6"
|
||||
},
|
||||
{
|
||||
"path": "docs/s2-g1-material-inventory-2026-11-14.md",
|
||||
"bytes": 5224,
|
||||
"sha256": "7a11927a3cb8424577a30a4a87643e78767f0f71874a5c8eea2f2c215bdf20a1"
|
||||
},
|
||||
{
|
||||
"path": "references/2026-11-14/s2-design/moldastra-note16-1900.md",
|
||||
"bytes": 19098,
|
||||
"sha256": "aa448d8e9b78a24ec925a61dacd30e35222792d07b1ced216b94758ab8bcf4fc"
|
||||
},
|
||||
{
|
||||
"path": "tools/inventory_g1_materials.py",
|
||||
"bytes": 7634,
|
||||
"sha256": "ab5fd9c5da55bd0f354e9b2dc4c24ccfcb05010b81ba6bceea0a71966651c64e"
|
||||
},
|
||||
{
|
||||
"path": "tools/simulate_g1_scheduler.py",
|
||||
"bytes": 6079,
|
||||
"sha256": "3395dbde8ce5966882936c54296d3fe06af4c4de9be34bc186c4eca1eda75dbf"
|
||||
},
|
||||
{
|
||||
"path": "tools/test_g1_scheduler_simulation.py",
|
||||
"bytes": 3588,
|
||||
"sha256": "245a3fb85418a39f8af8dc389fb736b7b0bb4b79f9c4576827ffd9125272a99e"
|
||||
}
|
||||
]
|
||||
}
|
||||
@@ -0,0 +1,36 @@
|
||||
{
|
||||
"scope": "existing material inventory only; no new descriptors, candidates, pair extraction, labels from original source, LOG reread, registration, training or GPU",
|
||||
"cases": 624,
|
||||
"source_keys": 249,
|
||||
"verified_parent_links": 0,
|
||||
"verified_independent_natural_positive_pairs": 0,
|
||||
"old_exposed_pairs": 30,
|
||||
"old_exposed_endpoint_cases": 56,
|
||||
"exposure_counts": {
|
||||
"UNKNOWN": 541,
|
||||
"KNOWN_DEVELOPMENT_EXPOSED": 83
|
||||
},
|
||||
"cluster_exposure_counts": {
|
||||
"EXPOSED_CLUSTER": 55,
|
||||
"UNKNOWN_EXPOSURE_CLUSTER": 194
|
||||
},
|
||||
"independent_holdout_eligibility_certified_cases": 0,
|
||||
"surface_cache_cases": 47,
|
||||
"surface_cache_records": 47,
|
||||
"npz_member_directories_inspected": 313,
|
||||
"legacy_geometry_metadata_cases": 624,
|
||||
"note16_complete_fingerprint_cases": 0,
|
||||
"custody": "IDENTITY_TABLES_PREPARED_LOCALLY; designated custodian/interface pending Claude; no anonymous IDs or blind bundle generated",
|
||||
"identity_manifest_sha256": "7a0bf4e91180a6745d18e0570510cec0744818ca5f2dd47a03e71388b96708d9",
|
||||
"input_sha256": {
|
||||
"artifacts/intake-2026-11-13/source_groups.json": "03ca9b1a30bd1b957a59dfae0561c452645f1a6151a72eb4da7cf1f51cc50e13",
|
||||
"artifacts/intake-2026-11-13/geometry_log_census.json": "29f107100ebcc296c41206434229b4941e573c9c4c8c57f820e82f388b2cf40b",
|
||||
"artifacts/intake-2026-11-13/geometry_screening.json": "98e0bee5bcfb4210ecf3c1c3554ef9c530b120f22c9fc3de5733a9619559937c"
|
||||
},
|
||||
"producer_sha256": "ab5fd9c5da55bd0f354e9b2dc4c24ccfcb05010b81ba6bceea0a71966651c64e",
|
||||
"interpretation": "0 certified independent positives/eligible cases means missing proof, not zero true relatives; unknown exposure cannot be promoted to unexposed. Old geometry screening is not the proposed fingerprint. Source keys are not CAD ancestry.",
|
||||
"mixed_metadata_disclosure": "Existing geometry_log_census JSON also stores historical process/output fields; parsed as a container, only bbox/covariance/point hash selected. Other fields not used or exported. NPZ access is ZIP member names only, no payload.",
|
||||
"cpu_seconds": 0.31049699999999997,
|
||||
"wall_seconds": 0.2523299700114876,
|
||||
"peak_rss_bytes": 21606400
|
||||
}
|
||||
@@ -0,0 +1,102 @@
|
||||
"""Pure discrete-event contract model. NO subprocesses, data reads or real jobs.
|
||||
|
||||
Virtual process-tree inputs test accounting and barriers, not Linux enforcement.
|
||||
A production adapter must supply sampled + final tree accounting, kill/reap,
|
||||
locks and persisted recovery; none is claimed here.
|
||||
"""
|
||||
from dataclasses import dataclass, asdict
|
||||
import math
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Task:
|
||||
pair_id: str
|
||||
rank: int
|
||||
duration: float = 1.
|
||||
tree_cpu_rates: tuple = (1.,)
|
||||
tree_rss_bytes: tuple = (1024,)
|
||||
disk_rate: float = 0.
|
||||
fails: bool = False
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Limits:
|
||||
workers: int = 16
|
||||
worker_cpu: float = 80*3600
|
||||
wall: float = 8*3600
|
||||
disk: float = 20*1024**3
|
||||
pair_wall: float = 1800.
|
||||
pair_rss: int = 1342177280
|
||||
checkpoint: int = 100
|
||||
|
||||
|
||||
def simulate(tasks, limits=Limits(), *, supervisor_cpu_rate=.1):
|
||||
"""Virtual seconds/bytes only. Each unique pair is attempted at most once."""
|
||||
tasks=sorted(tasks,key=lambda t:(t.rank,t.pair_id))
|
||||
if len({t.pair_id for t in tasks})!=len(tasks):raise ValueError('Duplicate pair cannot expand denominator')
|
||||
if not 1<=limits.workers<=16 or limits.checkpoint!=100:raise ValueError('Contract worker/checkpoint bound')
|
||||
if any(not math.isfinite(v) or v<=0 for v in asdict(limits).values()):raise ValueError('Invalid limit')
|
||||
if not math.isfinite(supervisor_cpu_rate) or supervisor_cpu_rate<0:raise ValueError('Invalid supervisor rate')
|
||||
for t in tasks:
|
||||
values=(t.duration,*t.tree_cpu_rates,*t.tree_rss_bytes,t.disk_rate)
|
||||
if not t.pair_id or not math.isfinite(t.rank) or any(not math.isfinite(v) or v<0 for v in values) or t.duration<=0 or not t.tree_cpu_rates or not t.tree_rss_bytes:
|
||||
raise ValueError('Invalid virtual tree/task')
|
||||
active={};records=[];events=[];checkpoints=[];index=0;end=min(100,len(tasks))
|
||||
elapsed=worker_cpu=supervisor_cpu=disk=0.;peak_workers=peak_tree_processes=0
|
||||
reason=None
|
||||
def finish(pid,why=None):
|
||||
a=active.pop(pid);t=a['task']
|
||||
records.append({'pair_id':pid,'rank':t.rank,'status':'FAIL' if why or t.fails else 'PASS',
|
||||
'reason':why or ('worker_exit' if t.fails else None),'worker_tree_cpu':a['cpu'],
|
||||
'tree_rss':sum(t.tree_rss_bytes),'start':a['start'],'finish':elapsed})
|
||||
events.append({'event':'finish','pair_id':pid,'time':elapsed,'reason':records[-1]['reason']})
|
||||
def global_limit():
|
||||
if worker_cpu>=limits.worker_cpu:return 'worker_cpu_cap'
|
||||
if elapsed>=limits.wall:return 'global_wall_cap'
|
||||
if disk>=limits.disk:return 'global_disk_cap'
|
||||
while index<len(tasks) or active:
|
||||
reason=global_limit()
|
||||
if reason:break
|
||||
# Barrier: never launch pair 101 while any of first 100 are in flight.
|
||||
while index<end and len(active)<limits.workers:
|
||||
t=tasks[index];index+=1
|
||||
active[t.pair_id]={'task':t,'start':elapsed,'cpu':0.}
|
||||
events.append({'event':'start','pair_id':t.pair_id,'time':elapsed,'attempt':index})
|
||||
peak_workers=max(peak_workers,len(active))
|
||||
peak_tree_processes=max(peak_tree_processes,sum(len(a['task'].tree_rss_bytes) for a in active.values()))
|
||||
if sum(t.tree_rss_bytes)>limits.pair_rss:finish(t.pair_id,'pair_tree_rss_cap')
|
||||
if not active and index==end:
|
||||
failed=sum(r['status']=='FAIL' for r in records)
|
||||
assert len(records)==index
|
||||
checkpoint={'attempted':index,'terminal':len(records),'failed':failed,'time':elapsed,'stop':failed*50>index}
|
||||
checkpoints.append(checkpoint);events.append({'event':'checkpoint',**checkpoint})
|
||||
if checkpoint['stop']:reason='failure_fraction_gt_2_percent';break
|
||||
if index==len(tasks):break
|
||||
end=min(end+100,len(tasks));continue
|
||||
rate=sum(sum(a['task'].tree_cpu_rates) for a in active.values())
|
||||
disk_rate=sum(a['task'].disk_rate for a in active.values())
|
||||
remaining=[min(a['task'].duration,limits.pair_wall)-(elapsed-a['start']) for a in active.values()]
|
||||
remaining += [limits.wall-elapsed]
|
||||
if rate:remaining.append((limits.worker_cpu-worker_cpu)/rate)
|
||||
if disk_rate:remaining.append((limits.disk-disk)/disk_rate)
|
||||
dt=max(0.,min(remaining))
|
||||
if dt==0:raise RuntimeError('Zero-time event without stop/terminal transition')
|
||||
elapsed+=dt;supervisor_cpu+=supervisor_cpu_rate*dt;disk+=disk_rate*dt
|
||||
for a in active.values():
|
||||
c=sum(a['task'].tree_cpu_rates)*dt;a['cpu']+=c;worker_cpu+=c
|
||||
for pid,a in list(active.items()):
|
||||
t=a['task'];age=elapsed-a['start']
|
||||
if age>=min(t.duration,limits.pair_wall):
|
||||
finish(pid,'pair_wall_cap' if t.duration>limits.pair_wall else None)
|
||||
if reason:
|
||||
for pid in list(active):finish(pid,reason)
|
||||
# Stop/global cap still records an end checkpoint against every attempted pair.
|
||||
if records and (not checkpoints or checkpoints[-1]['attempted']!=index):
|
||||
failed=sum(r['status']=='FAIL' for r in records)
|
||||
checkpoints.append({'attempted':index,'terminal':len(records),'failed':failed,'time':elapsed,'stop':failed*50>index,'trigger':'end/global stop'})
|
||||
assert len(records)==index and abs(sum(r['worker_tree_cpu'] for r in records)-worker_cpu)<1e-6
|
||||
return {'kind':'PURE_VIRTUAL_CONTRACT_SIMULATION_NOT_REAL_PROCESS_SUPERVISION',
|
||||
'status':'STOPPED' if reason else 'COMPLETE','stop_reason':reason,'attempted':index,
|
||||
'not_attempted':len(tasks)-index,'records':records,'events':events,'checkpoints':checkpoints,
|
||||
'virtual_worker_cpu':worker_cpu,'virtual_supervisor_cpu':supervisor_cpu,
|
||||
'virtual_wall':elapsed,'virtual_disk':disk,'peak_worker_trees':peak_workers,
|
||||
'peak_virtual_process_count_including_descendants':peak_tree_processes,
|
||||
'real_workers_started':0,'retry':False}
|
||||
@@ -0,0 +1,10 @@
|
||||
{
|
||||
"tests": 11,
|
||||
"passed": 10,
|
||||
"failed": 1,
|
||||
"failed_test": "test_final_short_batch_check",
|
||||
"error": "Expected failure_fraction_gt_2_percent for 1/49 failures, got None",
|
||||
"test_runner_seconds": 0.01,
|
||||
"cause": "Loop completed before final short-batch checkpoint could update stop reason",
|
||||
"source_sha256": "5b09c8e7288684cc07a2b789e0905ef3905a9030ea5778931a1b52212d8c4f21"
|
||||
}
|
||||
@@ -0,0 +1,12 @@
|
||||
{
|
||||
"returncode": 0,
|
||||
"wall_seconds": 0.08645775698823854,
|
||||
"test_process_cpu_seconds": 0.08585,
|
||||
"peak_rss_bytes": 14155776,
|
||||
"real_worker_processes": 0,
|
||||
"kind": "pure virtual contract simulation, not real OS/process-tree enforcement",
|
||||
"code_sha256": {
|
||||
"tools/simulate_g1_scheduler.py": "3395dbde8ce5966882936c54296d3fe06af4c4de9be34bc186c4eca1eda75dbf",
|
||||
"tools/test_g1_scheduler_simulation.py": "245a3fb85418a39f8af8dc389fb736b7b0bb4b79f9c4576827ffd9125272a99e"
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,16 @@
|
||||
test_child_rss_summed_and_failure_preserved (test_g1_scheduler_simulation.ContractTests) ... ok
|
||||
test_duplicate_pair_rejected (test_g1_scheduler_simulation.ContractTests) ... ok
|
||||
test_exact_two_percent_continues_and_barrier (test_g1_scheduler_simulation.ContractTests) ... ok
|
||||
test_final_short_batch_check (test_g1_scheduler_simulation.ContractTests) ... ok
|
||||
test_global_disk_cumulative (test_g1_scheduler_simulation.ContractTests) ... ok
|
||||
test_global_wall (test_g1_scheduler_simulation.ContractTests) ... ok
|
||||
test_global_worker_cpu_stops_all (test_g1_scheduler_simulation.ContractTests) ... ok
|
||||
test_pair_wall_failure_continues_to_checkpoint (test_g1_scheduler_simulation.ContractTests) ... ok
|
||||
test_stable_rank_and_worker_bound (test_g1_scheduler_simulation.ContractTests) ... ok
|
||||
test_three_percent_stops_no_101 (test_g1_scheduler_simulation.ContractTests) ... ok
|
||||
test_tree_cpu_and_failed_cpu_separate_from_supervisor (test_g1_scheduler_simulation.ContractTests) ... ok
|
||||
|
||||
----------------------------------------------------------------------
|
||||
Ran 11 tests in 0.012s
|
||||
|
||||
OK
|
||||
@@ -0,0 +1,42 @@
|
||||
# G1 现有材料盘点与模拟调度验收
|
||||
|
||||
本轮落实MoldAstra note16的19:00材料交接,不依赖k,不生成新候选、指纹、母件划分、盲包或配准。盘点的是现有记录的覆盖与出处,不能把缓存存在当作独立留验已通过。
|
||||
|
||||
| 项目 | 现有证据 |
|
||||
|---|---:|
|
||||
| 稳定件号 / 文件来源键 | 624 / 249 |
|
||||
| 旧已暴露候选对 / 不重复端点 | 30 / 56 |
|
||||
| 有明确个件开发暴露记录 | 83 |
|
||||
| 个件暴露史不能认证为未暴露 | 541 |
|
||||
| 含暴露成员的来源簇 / 暴露史未知来源簇 | 55 / 194 |
|
||||
| 有表面坐标与三角面缓存的件 | 47 |
|
||||
| 旧bbox/未加权点协方差/坐标哈希覆盖 | 624 |
|
||||
| note16完整面积/体积加权指纹覆盖 | 0 |
|
||||
| 有独立出处的母型关系 / 自然正对真值 | 0 / 0 |
|
||||
| 当前可认证为独立留验的件 | 0 |
|
||||
|
||||
最后三行的0表示现有审阅材料缺少所需证据,不能解释为没有共同母型或没有真实正对。文件来源键不被升级成CAD母型;旧30对由历史筛查器发现,没有独立自然正对依据,不充当留验。全部624件曾有自动粗几何筛查记录;83件的明确暴露标记来自旧30对端点与既有个件诊断/评分目录,其余只能标unknown。来源簇不拆开凑样本。逐件表保存字段缺口与证据路径,四类自然关系的可证独立母件数目前均缺,没有编造每类一个示例。
|
||||
|
||||
缓存盘点仅查看313个NPZ的ZIP成员名和文件stat,不解压数组、不重新核验几何有效性。47件缓存包含不同历史格式,含旧标签数组的容器已单独标记;它们不能直接发给盲审员。旧geometry_log_census是混合历史JSON,解析容器后仅取bbox、协方差、点坐标哈希三项;没有读取原始VTU/LOG,旧工况/分析输出字段不参与盘点或输出。旧描述符依赖网格节点分布,混合单元件未保证tet-only,不能替代note16新指纹,也不能用小缓存背景名次冻结全库k。
|
||||
|
||||
## 托管决定与实际封存
|
||||
|
||||
[Claude 19:05:35](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794654335163209)指定MeltAstra为托管人,MoldAstra为盲审员,Grok为揭盲见证。含身份文件已在Horsea的 `artifacts/s2-g1-inventory-2026-11-14/custodian/` 准备并封存,目录0700、文件0600。当前公共交付只有汇总、代码、验证和逐文件摘要,身份内容不进入盲审资料。此分离用于盲审流程;原记录完整保留,没有删除信息。原始盘点summary中的“待指定托管人”是生成时状态,已由 `public/custody-seal.json` 明确接续。
|
||||
|
||||
公开封存摘要包括parts.json、source_clusters.json、old_exposed_pairs.json、independent_natural_pairs.json与manifest.json。它们只是盘点身份表,不是已完成的匿名映射/抽样概率/留验答案;后者尚未产生。没有向审员发送任何身份表或盲包。局部权限不是加密或对同主机授权代理的强隔离,不作此类技术宣称。盲审标签先冻结提交,Grok核对已公布摘要,再按Claude流程对接映射。
|
||||
|
||||
19:05裁量还更新筛外层:任一路rank在k+1至3k为近层,rank>10k为远层;不沿用note16提议的2k层界。中间区间、两路分歧时如何构成互斥远层/完整抽样框尚需在抽样前明确,当前没有抽样。刚体/模板/共母型各类不足30/15/15的现状交Mold/Claude评定“预算不可验证”;不能因缺真值就把任意新对自动连边。
|
||||
|
||||
## 模拟调度验收:11项通过,真实并行监督尚未实现
|
||||
|
||||
`simulate_g1_scheduler.py`是纯离散事件合同模型,接收虚拟进程树CPU/RSS、耗时和磁盘增长,不启动任何worker、不打开研究数据。默认合同80 worker CPU-h、8h、20GiB、最多16个worker根;单对30min/1.25GiB,后代CPU/RSS纳入该worker树。监督CPU单列虚拟账,不从worker额度中混扣;实际监督预算仍须独立明确。
|
||||
|
||||
11项验收覆盖:严格2%边界(2/100继续、3/100停止)、第100对全部终态前禁止启动第101对、最终短批检查(1/49停止、1/50继续)、重复pair拒绝、失败与后代CPU计费、监督CPU分账、全局CPU/墙钟/累计磁盘停止、后代RSS合计、单对超时继续至检查点、稳定rank和并发根数上限。
|
||||
|
||||
首次11项中10项通过,短批最后一次检查未设置停止原因,1/49误报完成;源码快照和失败回执保留。修正结束检查后11/11通过,测试进程实测CPU0.085850秒、峰RSS13.5MiB。盘点CPU0.310497秒。测试中的8h/80h/字节数均为虚拟量,不能作为真实计算耗费。旧监督6项和旧配准7项没有重跑。
|
||||
|
||||
该模型没有真实Linux进程适配、全局锁、采样/最终wait4合账、kill/reap、故障恢复或生产队列启动能力。旧单队列监督也未被替换。因此“模拟规则通过”不等于真实16worker监督通过,更不授予启动80h科学队列。
|
||||
|
||||
## 下一交接
|
||||
|
||||
Mold根据汇总确认逐类证据缺口和是否可封包,Claude整合独立出处获取/开发样本条件。现有身份表由指定托管人保管;抽样框、种子、概率、盲包与完整指纹尚未生成。不重跑624LOG或旧10对,不训练、不使用GPU。
|
||||
@@ -0,0 +1,70 @@
|
||||
# S2 G1|跨来源近重复定标与冻结单(框架已采纳,数值待独立留验)
|
||||
|
||||
## 2026-11-14 17:38:03 裁决(上海时间)
|
||||
|
||||
**已采纳。** 同来源键强制同折。跨键确认的刚体近复制、镜像/缩放模板、共母型且流路有变化的版本分别编码,三类都同折防止泄漏;不由此推定流动等价。候选采用带物理尺度与尺度无关的双路召回。确认结合面积加权双向残差与小孔、薄桥核验,逐对记录 PASS、REJECT 或 UNRESOLVED。可能改变组件的未决对使 S2 保持 HOLD。既有 30 对按已暴露处理,旧 0.5%/1% 不移植。下文第 4 节的确定性分折及嵌套学习池取整规则已获采纳。[Claude 裁决原帖](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794649083662649)
|
||||
|
||||
**待实证、待授权。** 定标与留验须分开。独立留验后,才能冻结数值阈值、筛外抽查口径及漏筛/误连预算。校准回执、正式候选与证据边表、实际组件及 split ID 尚待形成。框架裁决本身不放行定标计算、十件测量、pilot 或训练;各阶段按 Yanqi 授权推进。
|
||||
|
||||
**下一阶段回顾。** 团队共同检查分组能支持哪一种“未见制件”主张,结合 Greif 的距离基线、Lee 的同族几何经验和实际留验证据再决定推进。G1 的整件同折规则与 O1 的件内分片继续分别评价。A1–A4 由 Claude 和 MeltAstra 维护。
|
||||
|
||||
## 18:51 接续|先交盲审清单,后定 k(MoldAstra 提议)
|
||||
|
||||
[Claude 的预算及顺序](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794653503976579) · [旧 10 对成本回执 108e7836](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/108e78364437c5f79304610df6c63dd4daad5e76/docs/s2-e1-cost-results-2026-11-14.md)。据固定回执,910.686 秒是修正版 worker CPU,监督另账。这 10 对来自已暴露候选。原任务回执明确独立留验未执行;目前核到的本地和 Drive 材料中也没有可发给盲审员的独立自然件对真值表。因此可预先规定描述符与取样方法,不能据此报已验证的 k 或距离阈值。
|
||||
|
||||
**请 MeltAstra 先交一份清单,而非启动新队列。** 用现有几何目录、缓存和已有来源证据列出可用母件、已知共同母型簇、已暴露 30 对及其端点、可独立于拟议筛选器取得的自然正对/易混淆对,以及是否已有可复用指纹;每项给证据路径和缺口。先用件号/来源键/已知母型关系隔离母件,再用预登记种子与稳定 ID 顺序划开发集与封存留验集。旧 30 对及已知相同母型不得进入独立留验。暴露范围单列;不同键只是检索对象,不能用作负例真值。历史工况、分析标签和模型结果均不进入盲包。
|
||||
|
||||
**最小交接件。** 两套材料分发:(a) 给盲审员的随机 pair ID、两件原单位表面及比例尺、统一渲染和孔/薄桥局部视图;隐藏来源键、候选来源、排名、残差、预算及折。审员先写两件是否有共同设计证据、关系码、局部依据或 UNRESOLVED;疑难交独立复核。(b) 托管清单保存 pair ID 对应的母件/来源、暴露标志、发现途径、母型溯源、入样分层与概率,留验答案在选择参数前封存。每个开发/留验包应覆盖刚体重剖、镜像或缩放模板、共母型改通路、近似外形但关键局部不同的歧义对。已知变换可作阳性合成探针,不能替代自然跨键同母型证据。先盘点每类独立母件数;缺自然正对时明确标缺,不能把每类一个示例当作已校准。涉及共同设计来源而仅凭几何无法定性的,由托管者在形状盲审后提供独立出处。
|
||||
|
||||
**描述符与选 k 的预案。** 沿用真 tet 外表面,建议形状路使用 log(A/R_g²)、log(V/R_g³)、体积加权中心二阶矩的有序特征值比(除以迹,保留两项)、面积加权表面半径/R_g 的 P10/P50/P90。物理路在相同向量上加 log(R_g/u),u 是固定原长度单位;各列定标用开发母件的中位数/IQR,IQR 为零的列按事前规则去除并登记,不能按留验结果选择。各路用标准化后的欧氏距离、含并列的双向前 k 邻居取并集。上述是待签的候选向量;沿用已定双路配准及 2,000 面积样点协议,不增粗配准过滤门。
|
||||
|
||||
开发时对独立来源证实的正对,保存每路每方向的伙伴名次。给定 k,四个名次中最小者 ≤k 就是候选召回。要求公布各关系类的独立母件数、命中/漏筛、未决数和留验确认误连;各特征、k 的选择准则、可接受错误及样本量由 Claude 在开留验答案前签字。一个可审的候选策略是取满足预设分关系类召回要求的最小 k,并用封存母件作一次接受/拒绝;留验失败即记版本失败,另定新留验,不能在同一留验上反复调 k。名次依赖干扰件数量:最终证据需在预定检索背景内(同一索引规则、排除同键/自身)给出;若现有指纹仅覆盖小缓存,只能做协议走查,不能用小库名次冻结 624 件的 k。可借既有预算内逐步形成的指纹表,对留验锚点做只读检索,无需提前输出正式候选并集或配准全库。
|
||||
|
||||
**筛外遗漏与放行。** 仅审前 k 返回的对无法估计漏筛。参数锁定后从筛外对抽查,区分近边缘与随机远对,保存各层总体数、抽样概率与审阅分母,不把人为挑选难例的通过率称作全库误率。若缺样本、预算或独立出处,报告“该关未验”;不得用旧 10 对或预算倒推 k。候选并集 >2,000 即报数,不降 k;独立留验、并集计数和资源审阅后再决定配准。Claude 的独立规划线是 ≤80 worker CPU-h、监督另账、≤16 workers、≤8 h、≤20 GiB;单对 ≤30 min/1.25 GiB,失败比例 >2% 停报,不能以单对样本的暖缓存外推作保证。此增补只指定材料交接,无全库操作。
|
||||
|
||||
## 18:55 接口合并|材料盘点、盲审和揭盲
|
||||
|
||||
本节按 [MeltAstra 的固定交接稿 07003020](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/07003020c5ea993a3afe0d5276743fcf486a308e/docs/s2-g1-calibration-handoff-2026-11-14.md) 与 [Claude 揭盲前的裁决](https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794653741166299) 更新前文状态。常规预算内准备工作可继续。以下“提议”仍需在抽样或开封前登记。
|
||||
|
||||
1. **母件与抽样框。** 下一份最小实物由 MeltAstra 从现有材料交一张清单:稳定件号、来源键、可证母型及依据、暴露或未知、现存表面/指纹路径及版本;已有件对另附两端、独立发现途径、关系证据。缺真实 ID、自然正对或可用指纹的栏写缺失及覆盖数。清单只向托管方交付。MoldAstra 据此给出逐类可用母件数、排除原因及是否足以封包。以来源键或已知母型相连的完整簇隔离开发与留验;暴露或暴露史不明的簇不计独立留验。提议用独立抽样种子 2026111455,对“协议版本、用途、种子、稳定簇/件对 ID”的 UTF-8 规范串做 SHA-256 排序,无放回抽样;先冻结抽样框哈希、分层、每层数量与跨层母族去重。最小盲包内容沿用 18:51 节。当前无可确认的独立母件分配表。
|
||||
2. **双路名次。** 向量、开发集缩放、检索背景、同键排除见 18:51 节。提议按未舍入的固定精度距离作竞争排名:单向 rank = 1 + 严格更近的合格件数。边的 rank 取两路×两方向之最小值,同时保留四个原名次;并列全收。k 仅在开发集选,封存集一次验收。未有完整背景和独立证据前 k 留空。
|
||||
3. **配准与错误预算。** 刚体原单位双向 d95、原尺寸差 |log(RgA/RgB)|、各自按 Rg 归一后的模板双向 d95 分栏;模板另存比例和镜像标志。开发集分别定通过带和复核带,记录单位、误差与局部证据。孔径/桥宽保存原尺寸、局部网格间距、剖面定位及两侧可解析性;测量带和分辨下限待开发探针验证。残差越界只排除对应配准分支,共母型改流路须另审来源。旧 ℓ/对角线数值不可转用。
|
||||
|
||||
Claude 已定独立留验自然正对按母族只计一次。单侧 95% 漏检率上界:刚体 ≤10% 且 ≥30 对,镜像/等比模板、共母型变体各 ≤20% 且各 ≥15 对。报告候选漏筛及最终漏连两级计数,未决单列;保守验收中未确认的正对计入最终未检出。开发集已连边另随机盲审 30 对,单侧 95% 误连率上界 ≤35%,此项只支持开发审计。建议正对采用精确二项上限,零漏时为 1−0.05^(1/n);同母族重复件不增加 n。材料不足的类标“预算不可验证”,按 Claude 的防泄漏倾向可另列保守同侧处理,其边不得充作已验证真值。灰区、实际样本量与抽样框、保守同侧的证据门槛由 Claude 在揭盲前签定。
|
||||
4. **筛外抽审。** Claude 定近边缘 200、随机远对 100,混入盲包。提议将筛外跨键无序对中 k < rank ≤ 2k 者归近层,其余归远层,先移除所有其他规则已召回的候选;层界须先签定。层内无放回,保存总体 N、实际 n、纳入概率 n/N;层不足则全取并登记缺额。分别报筛外发现阳性的比例及按总体加权的估计,该比例的分母是筛外件对,与自然正对召回的分母不同。约 5 人时仅覆盖 300 对各一分钟的首审,局部细查、争议复核与 30 对连边另计;超时先报告,依 Claude 裁决先减远层并登记新概率。
|
||||
5. **托管与开封。** Claude 在抽样前指定无调参职责的保管人。保管人生成随机匿名 ID,私存双向映射、种子、层/概率、来源证据与参考判定,公布盲包和映射文件的摘要值。审员仅得匿名几何与比例尺,交签名/版本化判定。MoldAstra 先提交开发选定的 k、分支带、实现版本与固定预测摘要;盲审和预测均锁定后,Claude 授权保管人按匿名 ID 对接,发布逐类计数、未决和偏离记录。证据不足或留验失败保留原记录。当前无保管人签收或开封记录。
|
||||
|
||||
**并行准备。** MeltAstra 可在清单盘点同时做合同与调度验收。按 Claude 已定规则,worker CPU 累计与监督 CPU 分账,整棵进程树守 16 workers、8 小时与 20 GiB 全局限额,逐对保留失败;失败比例为失败对数/已尝试对数,每 100 对及结束时检查,严格 >2% 停报。并发时检查点须对齐同一批已尝试件对的终态,重试同一对不得扩充分母。单对与全局预算沿用 18:51 节。原单队列首失败即停与该协议有差异,准备验收只用合同或模拟任务。本次交付限于纸面接口。
|
||||
|
||||
## 历史底稿|17:38 裁决前提交的原文
|
||||
|
||||
已读 [Claude v0.3 固定稿](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/69c58a77006325319903f7862054d461b00ca75c/docs/claude/s2-design-v0.3-2026-11-14.md) 和 [MeltAstra 执行差分 G1](https://gitea.frank14f.online/Frank14f/MeltMind/src/commit/33b18e08/docs/s2-execution-diff-v0-2026-11-14.md)。本页给出可先锁的程序与仍空缺的数值栏,没有重新生成候选、标注件对、测算预算或划折。
|
||||
|
||||
**估计目标。** 用户关心能否迁移到新的制件。Greif v5 的 624 件、249 个文件来源键为模拟样本与分组线索;同来源必须同侧,跨键仍可能复用设计。Lee 的同族新几何与跨设计来源回答不同问题。建议把 S2 主张限定为“在所声明的几何审查和同侧规则下留出的组件”。G1 判断的是两个整件能否分居训练/测试;O1 判断的是一个制件内部如何分片、如何沿真实接口交换量。两个几何可能需要同折却有不同流路,因此 G1 结果不授予流动等价性。拓扑/类型切分继续作为独立开放路线。[数据发布](https://zenodo.org/records/18598121) · [Lee 等](https://doi.org/10.1093/jcde/qwag026) · [前轮依据与近重复框架](https://drive.google.com/file/d/10aMWqeBFKuAmKgssaWv-qf-cteNQuVcz/view)
|
||||
|
||||
## 1 先锁输入、变换与同侧语义
|
||||
|
||||
- 只用真实 tet 子域、坐标单位、面连通关系与可靠 CAD 来源信息。节点号、网格密度、浇口、材料、过程日志、Fill_time 和误差均不作近重证据。全部同来源键已是同侧硬边;不因一个键含多个几何而拆开。精确内容哈希相同可快速确认,哈希不同不能排除重剖近重。
|
||||
- 推荐关系表:①平移及正规旋转下,同物理尺寸的同形/重剖近复制 → 同侧,记录 RIGID;②需要镜像或单一各向同性比例才吻合,且完整结构与局部复核通过 → 建议同侧,另记 MIRROR/SCALE_TEMPLATE、变换行列式与比例;③改变孔、薄桥或通道但有明确共用母型证据的版本 → 建议同侧,记 RELATED_VARIANT 并保存不同流路证据;仅有相似拓扑、厚度或局部筋型不足以连边。不得用自由形变把所有同类件“配准成功”。②③是防泄漏政策建议,需 Claude 在定标前接受或另定;与①分别计数。
|
||||
|
||||
## 2 宽筛 → 确认 → 三态审边
|
||||
|
||||
1. 每件记录体积 V、面积 A、体积中心及长度 ℓ=V^(1/3),尺寸保留原单位。低成本形状签名可用 A/ℓ²、按体积分的归一化惯性谱和面积加权的表面径向分位数(围绕体积中心、除以 ℓ);固定分位点、积分/面采样和舍入版本。按同一尺度/单位比较物理尺寸,同时开尺度无关通道检索模板;不能先以尺寸不合排除后者。对跨键描述符做组内特征尺度固定的近邻排序,取两个通道任一方向前 k 名之并集,加精确匹配及历史候选。平分距离不漏掉并列;优先级与处理顺序另按件号。k、特征尺度和可选距离带宽在校准后锁定。高相似仅作候选,不作为直接同侧证据。已看过的 30 对独立标记既有暴露,不能充当独立的盲测结果。
|
||||
2. 对候选做双向面积加权表面对面核验:由质心/惯性轴产生正规刚体起点,保留合规轴排列与符号,多起点局部细化;近重特征值时增加确定性的旋转覆盖或特征对应起点。事先锁旋转覆盖、初值上限、迭代上限与抽样精度。单一起点失败不得记 REJECT。报告最佳变换、收敛差异和残差图。d95 取两个方向“表面点到对方三角面距离”的面积加权 P95 中的较大者。刚体分支以成对长度 (ℓ_A+ℓ_B)/2 归一,同时核 |log(ℓ_A/ℓ_B)|;模板分支在各自归一尺度上配准,原尺度差独立记录。旧“对角线 0.5% / 尺寸差 1%”只作历史参照,分母已不同,数值不能搬用。
|
||||
3. 全局 P95 可漏掉覆盖不到 5% 面积的贯孔。按同一分辨率约定审阅双向残差尾部、窄孔开闭、薄桥及相邻但未连接的表面;只使用可解析尺度的证据。连接分量相同或惯性谱相似不等于流路相同。记录局部视图、定位和复核结论;边界对或关键拓扑差异由独立复核者盲于标签、折和误差审阅。
|
||||
4. 每对记录 PASS(关系代码及同侧证据)、REJECT(可排除该同侧关系的正面证据)或 UNRESOLVED(缺数据、配准退化、单位/桥孔不清、超预算、审阅分歧)。已确认“不是刚体近复制”不能自动排除模板关系。可能跨两个当前组件的 UNRESOLVED 使 G1 HOLD;已同组件的未决对保留审计,但不改变划折。筛外未审对也不冒充 REJECT。
|
||||
|
||||
## 3 定标证据与预算:先有独立检查,再封存数值
|
||||
|
||||
先固定定标集、留验集、样本抽取种子、关系政策、匹配指标和错误预算,之后才准调 k、配准与确认阈值。在定标集用刚体/重剖的已知同形对检验尺度与网格误差,用缩放、镜像、薄桥改动和同谱异形构造检验分支选择;另列自然件对的盲审证据。原 30 对只能帮助发现失效模式。保留不同母件的留验对;若重用了母件的多个变换,按母件统计,不把它们算成独立 CAD 家族。用筛外等概率或分层抽查(单列靠近筛选边缘与随机远对,记录抽样概率)查漏;刻意构造的负例不能替代筛外遗漏审计。筛查召回、误连及其通路风险分别记录。若留验对不足或结果不达预设误差预算,结论为“证据不足”,不能把未发现遗漏写成零遗漏。
|
||||
|
||||
**交 Claude 的冻结栏。** 必填:各分支 k/距离尺度、残差与尺寸的通过带/灰区、可解析通路尺度、留验母件及筛外复核数量、允许的漏筛/误连与升级规则;以及粗筛 CPU/内存、候选对总数、每对面采样和多起点 CPU/墙钟上限、盲审对数/人时上限。建议用“跨键描述符比较数×单次成本 + 候选数×每对配准上限 + 盲审人时”单列预算。实际候选或复核工作量超额则报告积压与影响,不按运行快慢丢件或把未审对记为非重复。当前没有验证这些栏的实测回执。阈值若用留验数据再改,需升版本并另立留验,不能沿用原验收;冻结规则后才生成正式候选清单。
|
||||
|
||||
## 4 确定性分组、划折与交接
|
||||
|
||||
确认边与同键边取并查集;组件 ID 由排序后的成员号生成。审查桥接边、最大组件和链式并入,保留正反证据,不将“同组件”解释为每对几何相似。建议把以下并列与取整规则交 Claude 锁定:组件按件数降序、同数按 seed 2026111409 派生的稳定组件顺序,依次放入当前部件总数最少的外折,平手先看组件数再看折号;记录各折件数、组件数和尺寸分布。每外折非测试组件以独立、按折编号派生的固定无标签顺序预留 ceil(20%·组件数) 作验证,其余组件用独立子流排序,学习池 n 个组件取 ceil(0.25n)、ceil(0.50n)、n 的同一前缀;公布实际数量。若不足五个可用组件、学习池为空或嵌套档位重合,记录不可形成该比较,不拆组件凑数。若 Claude 采用其他平衡目标,先把代价、并列顺序、固定 RNG 与排序版本写入契约。
|
||||
|
||||
最终交付应有:冻结的规则/版本及校准和留验回执;逐件几何和来源 manifest;带关系类型、状态、残差与局部图证的边表;组件—外折—验证—嵌套学习池表;development_exposed 清单(依据 S1 面板与既往诊断回执填实际件号,另记接触类型;无法确认的记录为 unknown,组件含暴露件时标记)。核验同键和同侧边零跨折,训练、内验证和测试组件互斥。没有上述实测交付前,本页不是正式 grouping key 或 split ID。
|
||||
|
||||
**本次状态。** 可审的是程序与口径;镜像/模板政策、数值阈值、预算、暴露件号全表和实际折均待定。此处不改变 G2–G5。G1 的定标执行及后续十件测量、pilot、全矩阵须遵循各自授权范围。
|
||||
|
||||
@@ -0,0 +1,79 @@
|
||||
"""Inventory existing metadata and NPZ member names only; no geometry processing.
|
||||
|
||||
Identity-bearing output is for the designated custodian, never a blind view.
|
||||
No original archive/VTU/LOG or NPZ payload is read. Historical mixed JSON is
|
||||
parsed, but only pre-existing geometry fields are selected; process/results
|
||||
fields are not exported or used. No parent identity is inferred from filenames.
|
||||
"""
|
||||
from pathlib import Path
|
||||
from zipfile import ZipFile
|
||||
from collections import Counter, defaultdict
|
||||
import hashlib,json,re,resource,time
|
||||
|
||||
CASE=re.compile(r'\d{8}_[a-zA-Z0-9]+_step_\d{3}')
|
||||
def write(path,data):
|
||||
path.write_text(json.dumps(data,indent=2,ensure_ascii=False,allow_nan=False)+'\n')
|
||||
def main():
|
||||
start=time.monotonic();out=Path('artifacts/s2-g1-inventory-2026-11-14');out.mkdir(exist_ok=False)
|
||||
private=out/'custodian';public=out/'public';private.mkdir();public.mkdir()
|
||||
base=Path('artifacts/intake-2026-11-13')
|
||||
inputs={n:base/n for n in ['source_groups.json','geometry_log_census.json','geometry_screening.json']}
|
||||
groups=json.loads(inputs['source_groups.json'].read_text());cases=sorted(c for v in groups.values() for c in v);case_set=set(cases)
|
||||
metadata=json.loads(inputs['geometry_log_census.json'].read_text())
|
||||
allowed=['bbox_size_raw','covariance_eigenvalues','exact_sorted_points_sha256']
|
||||
historic={r['case']:{k:r[k] for k in allowed if k in r} for r in metadata};del metadata
|
||||
screening=json.loads(inputs['geometry_screening.json'].read_text());pairs=screening['candidate_pairs'];assert len(pairs)==30
|
||||
exposed_old={p[k] for p in pairs for k in ['a','b']};assert exposed_old<=case_set
|
||||
caches=defaultdict(list);exposure=defaultdict(set);npz_count=0
|
||||
# Directory names are evidence of previous individual work, not independent truth.
|
||||
diagnostic_prefixes=('s1-real-scoring','s1-sampled-o2','e6-','o2-rep','o2a-','capacity-projectors','physical-support','routes-')
|
||||
for path in sorted(Path('artifacts').rglob('*')):
|
||||
if not path.is_file() or out in path.parents:continue
|
||||
ids=set(CASE.findall(str(path)))&case_set
|
||||
stage=path.parts[1]
|
||||
if stage.startswith(diagnostic_prefixes):
|
||||
for case in ids:exposure[case].add(stage)
|
||||
if path.suffix!='.npz':continue
|
||||
npz_count+=1
|
||||
with ZipFile(path) as z:keys=set(n.removesuffix('.npy') for n in z.namelist())
|
||||
surface=({'points','boundary'}<=keys or {'points_raw','boundary_triangles'}<=keys)
|
||||
if surface:
|
||||
st=path.stat();entry={'path':str(path),'bytes':st.st_size,'mtime_ns':st.st_mtime_ns,'surface_coordinate_key':'points' if 'points' in keys else 'points_raw','surface_index_key':'boundary' if 'boundary' in keys else 'boundary_triangles','contains_field_labels':'fill_time_raw' in keys,'validation':'ZIP member-name availability only; no arrays read or geometry validity asserted','producer_version':'E1 contract' if stage=='s2-e1-2026-11-14' else 'existing stage source/receipt; not revalidated'}
|
||||
for case in ids:caches[case].append(entry)
|
||||
rows=[]
|
||||
for case in cases:
|
||||
events=sorted(exposure[case]);known=case in exposed_old or bool(events)
|
||||
rows.append({'case':case,'source_key':case.split('_')[1],'source_key_evidence':str(inputs['source_groups.json']),
|
||||
'verified_parent_geometry':None,'parent_provenance':None,'parent_gap':'No independently verified CAD parent/source relationship found in existing reviewed manifests; source key is not a verified parent',
|
||||
'exposure_state':'KNOWN_DEVELOPMENT_EXPOSED' if known else 'UNKNOWN',
|
||||
'old_30_pair_endpoint':case in exposed_old,'individual_diagnostic_stages':events,
|
||||
'all_corpus_historical_screening':True,'unexposed_claim':False,'surface_caches':caches[case],
|
||||
'historical_geometry_metadata':historic.get(case,{}),'historical_fingerprint_source':str(inputs['geometry_log_census.json']),
|
||||
'historical_fingerprint_limit':'unweighted full point-cloud covariance/extent/hash; mesh-density dependent; mixed-cell cases not tet-only; not note16 area/volume weighted descriptor',
|
||||
'note16_complete_fingerprint':False,'note16_fingerprint_path':None,'independent_holdout_eligible':False})
|
||||
by_case={r['case']:r for r in rows};cluster_rows=[]
|
||||
for key,members in sorted(groups.items()):
|
||||
n=sum(by_case[c]['exposure_state']=='KNOWN_DEVELOPMENT_EXPOSED' for c in members)
|
||||
cluster_rows.append({'source_key':key,'members':sorted(members),'known_exposed_members':n,'state':'EXPOSED_CLUSTER' if n else 'UNKNOWN_EXPOSURE_CLUSTER','verified_parent_cluster':False,'holdout_eligible':False})
|
||||
old=[{'a':p['a'],'b':p['b'],'historical_rank':i+1,'discovery':'legacy covariance/extent nearest-neighbor screen','discovery_source':str(inputs['geometry_screening.json']),'development_exposed':True,'independent_discovery_evidence':None,'confirmed_relation':None,'natural_positive_truth':False} for i,p in enumerate(pairs)]
|
||||
write(private/'parts.json',rows);write(private/'source_clusters.json',cluster_rows);write(private/'old_exposed_pairs.json',old)
|
||||
write(private/'independent_natural_pairs.json',{'pairs':[],'status':'MISSING_INDEPENDENT_SOURCE_EVIDENCE','not_a_claim_of_absence_of_duplicates':True})
|
||||
private_manifest=[{'file':p.name,'sha256':hashlib.sha256(p.read_bytes()).hexdigest(),'bytes':p.stat().st_size} for p in sorted(private.glob('*.json'))]
|
||||
write(private/'manifest.json',private_manifest)
|
||||
summary={'scope':'existing material inventory only; no new descriptors, candidates, pair extraction, labels from original source, LOG reread, registration, training or GPU',
|
||||
'cases':len(rows),'source_keys':len(groups),'verified_parent_links':0,'verified_independent_natural_positive_pairs':0,
|
||||
'old_exposed_pairs':len(old),'old_exposed_endpoint_cases':len(exposed_old),
|
||||
'exposure_counts':dict(Counter(r['exposure_state'] for r in rows)),
|
||||
'cluster_exposure_counts':dict(Counter(r['state'] for r in cluster_rows)),
|
||||
'independent_holdout_eligibility_certified_cases':0,'surface_cache_cases':sum(bool(r['surface_caches']) for r in rows),
|
||||
'surface_cache_records':sum(len(r['surface_caches']) for r in rows),'npz_member_directories_inspected':npz_count,
|
||||
'legacy_geometry_metadata_cases':sum(bool(r['historical_geometry_metadata']) for r in rows),'note16_complete_fingerprint_cases':0,
|
||||
'custody':'IDENTITY_TABLES_PREPARED_LOCALLY; designated custodian/interface pending Claude; no anonymous IDs or blind bundle generated',
|
||||
'identity_manifest_sha256':hashlib.sha256((private/'manifest.json').read_bytes()).hexdigest(),
|
||||
'input_sha256':{str(p):hashlib.sha256(p.read_bytes()).hexdigest() for p in inputs.values()},
|
||||
'producer_sha256':hashlib.sha256(Path(__file__).read_bytes()).hexdigest(),
|
||||
'interpretation':'0 certified independent positives/eligible cases means missing proof, not zero true relatives; unknown exposure cannot be promoted to unexposed. Old geometry screening is not the proposed fingerprint. Source keys are not CAD ancestry.',
|
||||
'mixed_metadata_disclosure':'Existing geometry_log_census JSON also stores historical process/output fields; parsed as a container, only bbox/covariance/point hash selected. Other fields not used or exported. NPZ access is ZIP member names only, no payload.',
|
||||
'cpu_seconds':sum(resource.getrusage(resource.RUSAGE_SELF)[:2]),'wall_seconds':time.monotonic()-start,'peak_rss_bytes':resource.getrusage(resource.RUSAGE_SELF).ru_maxrss*1024}
|
||||
write(public/'summary.json',summary);print(json.dumps(summary,ensure_ascii=False,indent=2))
|
||||
if __name__=='__main__':main()
|
||||
@@ -0,0 +1,105 @@
|
||||
"""Pure discrete-event contract model. NO subprocesses, data reads or real jobs.
|
||||
|
||||
Virtual process-tree inputs test accounting and barriers, not Linux enforcement.
|
||||
A production adapter must supply sampled + final tree accounting, kill/reap,
|
||||
locks and persisted recovery; none is claimed here.
|
||||
"""
|
||||
from dataclasses import dataclass, asdict
|
||||
import math
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Task:
|
||||
pair_id: str
|
||||
rank: int
|
||||
duration: float = 1.
|
||||
tree_cpu_rates: tuple = (1.,)
|
||||
tree_rss_bytes: tuple = (1024,)
|
||||
disk_rate: float = 0.
|
||||
fails: bool = False
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Limits:
|
||||
workers: int = 16
|
||||
worker_cpu: float = 80*3600
|
||||
wall: float = 8*3600
|
||||
disk: float = 20*1024**3
|
||||
pair_wall: float = 1800.
|
||||
pair_rss: int = 1342177280
|
||||
checkpoint: int = 100
|
||||
|
||||
|
||||
def simulate(tasks, limits=Limits(), *, supervisor_cpu_rate=.1):
|
||||
"""Virtual seconds/bytes only. Each unique pair is attempted at most once."""
|
||||
tasks=sorted(tasks,key=lambda t:(t.rank,t.pair_id))
|
||||
if len({t.pair_id for t in tasks})!=len(tasks):raise ValueError('Duplicate pair cannot expand denominator')
|
||||
if not 1<=limits.workers<=16 or limits.checkpoint!=100:raise ValueError('Contract worker/checkpoint bound')
|
||||
if any(not math.isfinite(v) or v<=0 for v in asdict(limits).values()):raise ValueError('Invalid limit')
|
||||
if not math.isfinite(supervisor_cpu_rate) or supervisor_cpu_rate<0:raise ValueError('Invalid supervisor rate')
|
||||
for t in tasks:
|
||||
values=(t.duration,*t.tree_cpu_rates,*t.tree_rss_bytes,t.disk_rate)
|
||||
if not t.pair_id or not math.isfinite(t.rank) or any(not math.isfinite(v) or v<0 for v in values) or t.duration<=0 or not t.tree_cpu_rates or not t.tree_rss_bytes:
|
||||
raise ValueError('Invalid virtual tree/task')
|
||||
active={};records=[];events=[];checkpoints=[];index=0;end=min(100,len(tasks))
|
||||
elapsed=worker_cpu=supervisor_cpu=disk=0.;peak_workers=peak_tree_processes=0
|
||||
reason=None
|
||||
def finish(pid,why=None):
|
||||
a=active.pop(pid);t=a['task']
|
||||
records.append({'pair_id':pid,'rank':t.rank,'status':'FAIL' if why or t.fails else 'PASS',
|
||||
'reason':why or ('worker_exit' if t.fails else None),'worker_tree_cpu':a['cpu'],
|
||||
'tree_rss':sum(t.tree_rss_bytes),'start':a['start'],'finish':elapsed})
|
||||
events.append({'event':'finish','pair_id':pid,'time':elapsed,'reason':records[-1]['reason']})
|
||||
def global_limit():
|
||||
if worker_cpu>=limits.worker_cpu:return 'worker_cpu_cap'
|
||||
if elapsed>=limits.wall:return 'global_wall_cap'
|
||||
if disk>=limits.disk:return 'global_disk_cap'
|
||||
while index<len(tasks) or active:
|
||||
reason=global_limit()
|
||||
if reason:break
|
||||
# Barrier: never launch pair 101 while any of first 100 are in flight.
|
||||
while index<end and len(active)<limits.workers:
|
||||
t=tasks[index];index+=1
|
||||
active[t.pair_id]={'task':t,'start':elapsed,'cpu':0.}
|
||||
events.append({'event':'start','pair_id':t.pair_id,'time':elapsed,'attempt':index})
|
||||
peak_workers=max(peak_workers,len(active))
|
||||
peak_tree_processes=max(peak_tree_processes,sum(len(a['task'].tree_rss_bytes) for a in active.values()))
|
||||
if sum(t.tree_rss_bytes)>limits.pair_rss:finish(t.pair_id,'pair_tree_rss_cap')
|
||||
if not active and index==end:
|
||||
failed=sum(r['status']=='FAIL' for r in records)
|
||||
assert len(records)==index
|
||||
checkpoint={'attempted':index,'terminal':len(records),'failed':failed,'time':elapsed,'stop':failed*50>index}
|
||||
checkpoints.append(checkpoint);events.append({'event':'checkpoint',**checkpoint})
|
||||
if checkpoint['stop']:reason='failure_fraction_gt_2_percent';break
|
||||
if index==len(tasks):break
|
||||
end=min(end+100,len(tasks));continue
|
||||
rate=sum(sum(a['task'].tree_cpu_rates) for a in active.values())
|
||||
disk_rate=sum(a['task'].disk_rate for a in active.values())
|
||||
remaining=[min(a['task'].duration,limits.pair_wall)-(elapsed-a['start']) for a in active.values()]
|
||||
remaining += [limits.wall-elapsed]
|
||||
if rate:remaining.append((limits.worker_cpu-worker_cpu)/rate)
|
||||
if disk_rate:remaining.append((limits.disk-disk)/disk_rate)
|
||||
dt=max(0.,min(remaining))
|
||||
if dt==0:raise RuntimeError('Zero-time event without stop/terminal transition')
|
||||
elapsed+=dt;supervisor_cpu+=supervisor_cpu_rate*dt;disk+=disk_rate*dt
|
||||
for a in active.values():
|
||||
c=sum(a['task'].tree_cpu_rates)*dt;a['cpu']+=c;worker_cpu+=c
|
||||
for pid,a in list(active.items()):
|
||||
t=a['task'];age=elapsed-a['start']
|
||||
if age>=min(t.duration,limits.pair_wall):
|
||||
finish(pid,'pair_wall_cap' if t.duration>limits.pair_wall else None)
|
||||
reason=reason or global_limit()
|
||||
if reason:
|
||||
for pid in list(active):finish(pid,reason)
|
||||
# Stop/global cap still records an end checkpoint against every attempted pair.
|
||||
if records and (not checkpoints or checkpoints[-1]['attempted']!=index):
|
||||
failed=sum(r['status']=='FAIL' for r in records)
|
||||
checkpoints.append({'attempted':index,'terminal':len(records),'failed':failed,'time':elapsed,'stop':failed*50>index,'trigger':'end/global stop'})
|
||||
if reason is None and checkpoints and checkpoints[-1]['stop']:
|
||||
reason='failure_fraction_gt_2_percent'
|
||||
assert len(records)==index and abs(sum(r['worker_tree_cpu'] for r in records)-worker_cpu)<1e-6
|
||||
return {'kind':'PURE_VIRTUAL_CONTRACT_SIMULATION_NOT_REAL_PROCESS_SUPERVISION',
|
||||
'status':'STOPPED' if reason else 'COMPLETE','stop_reason':reason,'attempted':index,
|
||||
'not_attempted':len(tasks)-index,'records':records,'events':events,'checkpoints':checkpoints,
|
||||
'virtual_worker_cpu':worker_cpu,'virtual_supervisor_cpu':supervisor_cpu,
|
||||
'virtual_wall':elapsed,'virtual_disk':disk,'peak_worker_trees':peak_workers,
|
||||
'peak_virtual_process_count_including_descendants':peak_tree_processes,
|
||||
'real_workers_started':0,'retry':False}
|
||||
@@ -0,0 +1,55 @@
|
||||
import unittest
|
||||
from dataclasses import replace
|
||||
from simulate_g1_scheduler import Task, Limits, simulate
|
||||
|
||||
|
||||
def tasks(n, failures=()):
|
||||
return [Task(f'sim-{i:04}',i,fails=i in failures) for i in range(n)]
|
||||
|
||||
class ContractTests(unittest.TestCase):
|
||||
def test_exact_two_percent_continues_and_barrier(self):
|
||||
ts=tasks(201,{0,1});ts[0]=replace(ts[0],duration=20)
|
||||
r=simulate(ts)
|
||||
self.assertEqual(r['status'],'COMPLETE');self.assertEqual([c['attempted'] for c in r['checkpoints']],[100,200,201])
|
||||
first_end=max(x['finish'] for x in r['records'] if x['rank']<100)
|
||||
self.assertGreaterEqual(next(x['start'] for x in r['records'] if x['rank']==100),first_end)
|
||||
self.assertEqual(r['peak_worker_trees'],16)
|
||||
def test_three_percent_stops_no_101(self):
|
||||
r=simulate(tasks(250,{0,1,2}));self.assertEqual(r['attempted'],100);self.assertEqual(r['not_attempted'],150)
|
||||
self.assertEqual(r['stop_reason'],'failure_fraction_gt_2_percent')
|
||||
def test_final_short_batch_check(self):
|
||||
self.assertEqual(simulate(tasks(49,{0}))['stop_reason'],'failure_fraction_gt_2_percent')
|
||||
self.assertEqual(simulate(tasks(50,{0}))['status'],'COMPLETE')
|
||||
self.assertEqual(simulate(tasks(100,{0,1,2}))['stop_reason'],'failure_fraction_gt_2_percent')
|
||||
def test_duplicate_pair_rejected(self):
|
||||
with self.assertRaises(ValueError):simulate([Task('same',1),Task('same',2)])
|
||||
def test_tree_cpu_and_failed_cpu_separate_from_supervisor(self):
|
||||
r=simulate([Task('a',0,2,(1.,2.),(10,20),fails=True)],supervisor_cpu_rate=5)
|
||||
self.assertEqual(r['virtual_worker_cpu'],6);self.assertEqual(r['virtual_supervisor_cpu'],10)
|
||||
self.assertEqual(r['records'][0]['worker_tree_cpu'],6)
|
||||
def test_global_worker_cpu_stops_all(self):
|
||||
r=simulate([Task('a',0,10),Task('b',1,10)],replace(Limits(),worker_cpu=3))
|
||||
self.assertEqual(r['stop_reason'],'worker_cpu_cap');self.assertEqual(r['virtual_worker_cpu'],3)
|
||||
self.assertEqual(len(r['records']),2);self.assertTrue(all(x['status']=='FAIL' for x in r['records']))
|
||||
def test_global_wall(self):
|
||||
r=simulate([Task('a',0,10)],replace(Limits(),wall=2))
|
||||
self.assertEqual(r['stop_reason'],'global_wall_cap');self.assertEqual(r['virtual_wall'],2)
|
||||
def test_global_disk_cumulative(self):
|
||||
r=simulate([Task('a',0,1,disk_rate=6),Task('b',1,10,disk_rate=4)],replace(Limits(),disk=14))
|
||||
self.assertEqual(r['stop_reason'],'global_disk_cap');self.assertEqual(r['virtual_disk'],14)
|
||||
def test_child_rss_summed_and_failure_preserved(self):
|
||||
ts=tasks(100);ts[0]=Task('sim-0000',0,1,(1,1),(70,70))
|
||||
r=simulate(ts,replace(Limits(),pair_rss=100000))
|
||||
self.assertEqual(r['records'][0]['tree_rss'],140)
|
||||
r=simulate([Task('x',0,1,(1,1),(70,70))],replace(Limits(),pair_rss=100))
|
||||
self.assertEqual(r['records'][0]['reason'],'pair_tree_rss_cap')
|
||||
def test_pair_wall_failure_continues_to_checkpoint(self):
|
||||
ts=tasks(100);ts[0]=replace(ts[0],duration=5)
|
||||
r=simulate(ts,replace(Limits(),pair_wall=2))
|
||||
self.assertEqual(r['status'],'COMPLETE');self.assertEqual(next(x for x in r['records'] if x['rank']==0)['reason'],'pair_wall_cap')
|
||||
def test_stable_rank_and_worker_bound(self):
|
||||
r=simulate([Task('z',1),Task('b',0),Task('a',0)],replace(Limits(),workers=1))
|
||||
self.assertEqual([e['pair_id'] for e in r['events'] if e['event']=='start'],['a','b','z'])
|
||||
with self.assertRaises(ValueError):simulate(tasks(2),replace(Limits(),workers=17))
|
||||
|
||||
if __name__=='__main__':unittest.main()
|
||||
Reference in new issue
Block a user