kali20gakki/msAgent / skills/accuracy/loss-gnorm-diff-analyzer
loss-gnorm-diff-analyzer
Verified端到端分析两侧 msprobe PyTorch statistics 原始 dump 数据(dump.json + stack.json + construct.json)与精度问题现象(loss / grad norm / 输出不一致等),按「确定性预处理 → 多视角候选发现 → 汇总排序 → 分级定向验证 → 报告校验」流程识别可疑候选节点并构建到目标异常现象的因果链,交付主分析报告与结构化证据记录。当用户提供两侧 dump 目录并需要定位训练精度问题根因时使用。侧别判定:用户指定 target(待分析侧)/ golden(参考侧)优先;未指定时按目录名与数据内容推断(npu/gpu 等仅为软线索,不作硬性假设);不可靠时询问用户。
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 31.9 KB |
| references/alignment-playbook.md | markdown | 4.8 KB |
| references/candidate-discovery.md | markdown | 6.1 KB |
| references/causal-validation.md | markdown | 4.8 KB |
| references/dump-analysis-mechanism.md | markdown | 44.9 KB |
| references/evidence-and-state.md | markdown | 11.2 KB |
| references/hard-constraints.md | markdown | 68.6 KB |
| references/report-validation.md | markdown | 12.4 KB |
| scripts/dump_analyzer.py | script | 77.4 KB |
| scripts/make_evidence.py | script | 5.7 KB |
| scripts/query.py | script | 63.4 KB |
| scripts/report_check.py | script | 8.5 KB |
scripts/make_evidence.py
View on GitHub#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
make_evidence.py — evidence.json 骨架 / 结构校验。
消灭末段机械修复循环的两类来源:
1. 结构缺字段 → `new` 落的骨架自带 report_check.py 要求的全部必填顶层
字段,先天结构合规;
2. 末段才发现的结构问题 → `check` 随时可跑(规则=report_check.py 的
E1–E10;--report 给出时含 R1–R4/E9),填充过程中即校验即修复。
资格表(candidates[].qualification)与排除/锚点结案台账(exclusion_ledger)
随骨架/形状契约一并落 evidence.json——报告不承载机器可查结构。
用法(python3 -I 运行):
new --case caseX --out <工作目录>/evidence.json
分析开工第一件事。骨架中候选/证据为空,此时跑 check 会报
E4(candidates 为空)——属预期草稿态,填入后消除。
check --evidence <路径> [--report <report.md>]
只校验不改文件;交付前最后一步也用它。
"""
import argparse
import json
import os
import sys
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
import report_check # noqa: E402 (同目录,规则唯一权威,不复制实现)
# 形状契约:new 时打印一次,条目照抄该形状填充(消灭首稿结构违规返工)。
# 枚举与 report_check.py 常量同源(CANDIDATE_KEYS/CHAIN_STATUS/ROOT_STATUS)。
SHAPE_NOTICE = """ 条目形状契约(照抄填充,勿自创结构):
candidates[] 条目 = {"candidate_id", "node", "rank", "basis",
"supporting_evidence_ids": [证据ID...], "counterevidence_ids": [...],
"causal_chain": [{"desc": str, "status": 枚举, "evidence_ids": [...]}],
"confidence", "role", "qualification"}(链环节必须是含 status 的 dict,
不能是裸字符串)
qualification = HC-C8 资格表行,四格:
{"predicate_rewrite_test", "anchor_verification",
"background_unexcluded", "verdict"}
(锚点验证格合并承载:生产者锚点/节点级验证/同缺陷类归一;背景级与
未排除格含 HC-C5 迁移留痕;资格结论格含机制覆盖内/外;schema 细节
见 evidence-and-state.md)
exclusion_ledger[] 条目 = {"item", "type", "disposition", "basis"}
(排除/降权项与速览锚点结案台账;同类条目合并为族级行,参考上限 5)
evidence_log[] 条目 = {"id", "source", "metrics", "usage"}
(usage=用途与验证结论一句;不设独立 verification 字段)
root_cause_judgment = {"status", "statement", "confidence",
"unverified_links"}
枚举表(填其它值 report_check 会打回):
causal_chain[].status ∈ {已证实, 合理推断, 待验证}
root_cause_judgment.status ∈ {已验证根因, 较强候选, 待验证假设}
(位序/限定语一律写进 statement,不进 status)"""
SKELETON = {
"case": None,
"phenomenon": {"description": "", "evidence_sources": [],
"quantified": ""},
"candidates": [],
"evidence_log": [],
"exclusion_ledger": [],
"open_questions": [],
"root_cause_judgment": {"status": "", "statement": "", "confidence": "",
"unverified_links": []},
"verification_methods_used": [],
"run_notes": "",
}
def die(msg, code=2):
print("[make_evidence.py] 错误: %s" % msg, file=sys.stderr)
sys.exit(code)
def load_evidence(path):
try:
with open(path, "r", encoding="utf-8") as f:
ev = json.load(f)
except OSError as e:
die("无法读取 %s: %s" % (path, e))
except json.JSONDecodeError as e:
die("%s 不是合法 JSON: %s" % (path, e))
if not isinstance(ev, dict):
die("%s 顶层不是对象" % path)
return ev
def save(path, ev):
with open(path, "w", encoding="utf-8") as f:
json.dump(ev, f, ensure_ascii=False, indent=2)
def cmd_new(args):
if os.path.exists(args.out):
die("%s 已存在——不覆盖(防重置已填内容);如需重建请先删除"
% args.out)
ev = json.loads(json.dumps(SKELETON))
ev["case"] = args.case
save(args.out, ev)
print("make_evidence new: 骨架已写入 %s" % args.out)
print(" 待填:phenomenon / candidates / evidence_log / "
"exclusion_ledger / open_questions / root_cause_judgment / "
"verification_methods_used")
print(SHAPE_NOTICE)
print(" 交付前最后一步:check --evidence %s --report <report.md>"
% args.out)
def run_check(evidence_path, report_path):
issues = []
n_cands = None
if report_path:
n_cands = report_check.check_report(report_path, issues)
report_check.check_evidence(evidence_path, issues, n_cands)
if issues:
print("共 %d 项问题:" % len(issues))
for it in issues:
print(" - %s" % it)
return 1
print("make_evidence check: 通过(结构完整性;语义自检仍需人工完成)")
return 0
def cmd_check(args):
sys.exit(run_check(args.evidence, args.report))
def build_parser():
p = argparse.ArgumentParser(prog="make_evidence.py",
description="evidence.json 骨架/结构校验")
sub = p.add_subparsers(dest="cmd", required=True)
s = sub.add_parser("new")
s.add_argument("--case", required=True)
s.add_argument("--out", required=True)
s.set_defaults(func=cmd_new)
s = sub.add_parser("check")
s.add_argument("--evidence", required=True)
s.add_argument("--report")
s.set_defaults(func=cmd_check)
return p
def main():
args = build_parser().parse_args()
args.func(args)
if __name__ == "__main__":
main()
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |