Files
MeltMind/tools/report_r4bc_final.py

123 lines
10 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""Read frozen results and receipts only; no predictions, metrics or bootstrap rerun."""
import collections
import datetime
import math
from pathlib import Path
import time
from zoneinfo import ZoneInfo
import r4bc_final as f
import r4bc_determinism as d
from r4bc_common import read,save,sha,checked
def main():
start=time.process_time();f.pins();d.pins()
root=f.ROOT;det=d.ROOT
exact=read(root/'INTERIM_100-exact-comparison.json');bits=read(det/'bitwise-comparison.json');end=read(det/'complete.json')
assert exact['status']=='EXACT_MATCH' and not exact['defects']
assert bits['status']=='BITWISE_MATCH' and not bits['defects'] and bits['models']==35 and bits['prediction_arrays']==3864
assert end['status']=='COMPLETE_BITWISE_MATCH'
summary=read(root/'scores/summary.json');rows=read(checked(root/'scores/part-scores.json',summary['part_scores_sha256']))
frozen=read(checked(root/'predictions-freeze.json',summary['predictions_freeze_sha256']))
keys=lambda xs:{(x['pool'],x['case'],x['arm']) for x in xs}
assert len(rows)==len(frozen['entries'])==14904 and len(keys(rows))==14904 and keys(rows)==keys(frozen['entries'])
assert not frozen['missing']
paths=list((root/'prediction-guards').glob('*/worker.receipt.json'))+[root/'score-guard/worker.receipt.json',root/'compare-guard/worker.receipt.json']
paths+=list((det/'prediction-guards').glob('*/worker.receipt.json'))+[det/'compare-guard/worker.receipt.json']
receipts=[read(p) for p in paths]
assert len(receipts)==108 and all(x['status']=='PASS' and not x['monitor_errors'] and not x['retry'] and not x['gpu_visible'] for x in receipts)
for x in receipts:
assert x['cpu_seconds']<=x['limits']['cpu_seconds'] and x['wall_seconds']<=x['limits']['wall_seconds']
assert x['peak_tree_rss_bytes']<=x['limits']['rss_bytes'] and x['peak_process_count']<=1
metrics=['relative_l2','kendall_tau_b','last_location','near_wall']
undefined={m:sum(r[m]['value'] is None or not math.isfinite(r[m]['value']) for r in rows) for m in metrics}
means={};contrasts={};domains={};boot_undefined=0
for pool,p in summary['pools'].items():
assert p['complete_five_fold_all_arms']
means[pool]={};domains[pool]={}
for metric,m in p['metrics'].items():
means[pool][metric]={**m['primary']['means'],**m['chain_common']['means'],'H0':m['edges']['H0->H0p']['means']['H0']}
domains[pool][metric]=sorted({(z['parts'],z['components']) for z in [m['primary'],m['chain_common'],*m['edges'].values()]})
primary=p['metrics']['relative_l2']['primary']['differences']
edges={k:v['difference'] for k,v in p['metrics']['relative_l2']['edges'].items()}
contrasts[pool]={k:{field:v[field] for field in ['value','low','high','components','undefined_replicates']} for k,v in {**primary,**edges}.items()}
boot_undefined+=sum(x['undefined_replicates'] for x in contrasts[pool].values())
original=read(f.ORIGINAL/'batch-complete.json');post_cpu=end['post_cpu_seconds']
deadline=read(root/'adapter-freeze.json')['shared_post_start_unix']+21600
finished=(det/'complete.json').stat().st_mtime
assert finished<=deadline and post_cpu<=14400 and post_cpu+original['fit_cpu_seconds']+original['setup_cpu_seconds']<=144000
audit=dict(status='PASS',domains=domains,undefined_metrics=undefined,undefined_bootstrap=boot_undefined,
scoring_rows=len(rows),prediction_entries=len(frozen['entries']),checked_new_guard_receipts=len(receipts),
exact_interim=exact,bitwise_status=bits['status'],bitwise_arrays=3864,
scientific_source_pins='PASS',no_new_scoring_predictions_or_decodes_in_audit=True,
post_cpu_seconds=post_cpu,total_worker_cpu_seconds=post_cpu+original['fit_cpu_seconds']+original['setup_cpu_seconds'],
finished_unix=finished,shared_deadline_unix=deadline,all_new_resource_guards_pass=True,
peak_new_worker_rss_bytes=max(x['peak_tree_rss_bytes'] for x in receipts),
source_sha256={str(p):sha(p) for p in [root/'scores/summary.json',root/'scores/part-scores.json',root/'predictions-freeze.json',root/'INTERIM_100-exact-comparison.json',det/'bitwise-comparison.json',det/'complete.json']},
metadata_cpu_seconds_within_reserved_budget=time.process_time()-start)
save(root/'final-result-audit.json',audit);save(root/'final-result-tables.json',dict(means=means,contrasts=contrasts,domains=domains,undefined_metrics=undefined))
fmt=lambda x:f'{x:.12f}'
arms=['EIK','H0','H0p','H1','H2','H3','H4','A','Q']
finish=datetime.datetime.fromtimestamp(finished,ZoneInfo('Asia/Shanghai')).isoformat()
report=f'''# 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{finish},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|
|---|---:|---:|---:|
'''
for arm in arms:report+='|'+arm+'|'+'|'.join(fmt(means[p]['relative_l2'][arm]) for p in ['100','50','25'])+'|\n'
report+='''
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|
|---|---|---:|---|
'''
for pool in ['100','50','25']:
for name,v in contrasts[pool].items():report+=f"|{pool}%|{name}|{fmt(v['value'])}|[{fmt(v['low'])}, {fmt(v['high'])}]|\n"
report+='''
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|
|---|---|---:|---:|---:|
'''
for pool in ['100','50','25']:
for arm in arms:report+=f'|{pool}%|{arm}|'+'|'.join(fmt(means[pool][m][arm]) for m in ['kendall_tau_b','last_location','near_wall'])+'|\n'
report+=f'''
## 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{undefined};27个主/相邻比较的未定义bootstrap总数{boot_undefined}。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{post_cpu:.6f}/14400CPU秒(含原interim346.870387、新合成与90秒元数据保守预留),总worker累计{audit['total_worker_cpu_seconds']:.6f}/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{audit['metadata_cpu_seconds_within_reserved_budget']:.6f}CPU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。
'''
(f.BASE/'docs/r4bc-final-results-2026-11-17.md').write_text(report)
print({k:v for k,v in audit.items() if k not in ['domains','source_sha256','exact_interim']})
if __name__=='__main__':main()