AT-004 · archived verified run

评测预处理漂移：已隔离定位并可信修复

六个 Agent 基于归档证据完成诊断、受控实验与独立复核。这里展示的是经过完整性校验的 Evidence Replay，不是实时运行界面。

PASS / RESOLVED6 Agent roles ranHuman approval before executionTrace CHAIN_OK

Incident Summary

PUBLIC ALLOWLIST VIEW

Archived task

LABOPS-AT-004-EVAL-DRIFT

Evaluation preprocessing drift

Decision

PASS

Independent verification

Resolution

RESOLVED

Target ≥ 97.00%

Evidence

10 facts

27 ZIP entries

Before / After Metrics

THREE REPEATS · ZERO SPREAD

Before · train_augmented

71.88%

71.88% × 3

After · eval_standard

97.81%

97.81% × 3

六 Agent 协作顺序

HUMAN APPROVAL IS A SEPARATE GATE

Incident CommanderRAN

Evidence CollectorRAN

RCA AnalystRAN

Experiment PlannerRAN

Safe ExecutorRAN

Verification AuditorRAN

Commander编排与封包，不能覆盖终态裁决

Collector只能取证，不能诊断

Analyst只能诊断，不能执行

Planner只能制定计划，不能审批

Executor只能执行获批计划，不能宣布成功

Auditor独占终态裁决，不能修改结果

Evidence 排除过程

HASHES FIRST · CAUSATION BY CONTROLLED EXPERIMENT

EXCLUDED

Checkpoint

Current hash equals the archived reference hash.

EXCLUDED

Validation data

Dataset hash equals the archived reference hash.

EXCLUDED

metric.py

Metric implementation hash is unchanged.

EXCLUDED

Random variance

Three baseline repeats have zero spread.

CONFIRMED

Preprocessing drift

The controlled one-variable experiment restored the target metric.

Experiment Plan

只改变一个评估配置

eval_config.json · evaluation.preprocessing_profile

train_augmented → eval_standard

BudgetCPU · no network · ≤ 30s · 3 repeats

Approval gateExplicit and required

RollbackRestore the original profile or discard the sandbox

Forbiddenmetric.py, validation_data.pt, checkpoint, evaluation_protocol.yaml, thresholds, target_metric, original_workspace

Human Approval

APPROVED

人工审批是独立安全门，不计入六个 Agent 角色。

Approval IDLABOPS-AT-004-APPROVAL-001

Approved at2026-08-04T17:06:27.191Z

Before executionYES

ScopeOnly the allowlisted one-variable sandbox experiment

Runner Execution Result

PASS

Archived imagelabops/pytorch-cpu-runner:0.2.0

Commandevaluate_preprocessing_profile

IsolationCPU-only · no network · sandbox-only

RuntimePython 3.11.15 · torch 2.5.1+cpu

Protected hashesVERIFIED

Artifact hashesVERIFIED

Runtime Capability Check

All preflight controls passed

runner image

torch

checkpoint

config

paths

resource budget

command allowlist

plan policy

Trace

CHAIN OK, 7 ENTRIES · FINAL CHAIN_OK / ACCEPTED

1/6

labops-manager (Incident Commander) → evidence-collector

task accepted dispatched · 2026-08-04T16:52:40Z

accepted/dispatched

2/6

evidence-collector → rca-analyst

hypotheses ranked · 2026-08-04T16:59:45Z

DONE

3/6

rca-analyst → experiment-planner

plan approved policy · 2026-08-04T17:04:45Z

AUTO_APPROVED

4/6

experiment-planner → safe-executor

run executed · 2026-08-04T17:11:32Z

EXECUTION_COMPLETED

5/6

safe-executor → verification-auditor

verification completed · 2026-08-04T17:13:50Z

RESOLVED

6/6

verification-auditor → labops-manager (final packaging)

final audit accepted · 2026-08-04T17:42:00Z

CHAIN_OK/ACCEPTED

Reusable Infra

NOT A ONE-OFF SCRIPT

可复用能力合同

7 versioned Skills

同一组职责边界可迁移到训练、评测、数据处理和发布验证。

Structured I/O SchemaSandbox execution contractPolicy + approval gateCase Memory

Auditor Decision

INDEPENDENT RECOMPUTATION

Verification Auditor

PASS / RESOLVED

三次候选结果达到冻结阈值；重复性、审批顺序、受保护输入、Runner 产物和 Trace 哈希链均通过独立复核。

VERIFIED

Evidence Bundle SHA-256

27 ZIP ENTRIES · MEMBER SET VERIFIED

AT-004 archived evidence bundle · 27 ZIP entries

4092b43f39df52db3847caa28ca01e4321129a1c17ec7ca5efd2029ab1fb77cd

HASH VERIFIED

辅助安全分支

ARCHIVED CASES · UNCHANGED

AT-002 · BLOCKED

缺少运行依赖时安全阻塞

The approved checkpoint experiment stayed blocked because the Worker evaluation environment did not provide torch; no passing postcondition was claimed.

Valid caseINCONCLUSIVE

ResolutionDEMO_PASSED_NOT_RESOLVED

POLICY_VIOLATION / ROLLED_BACK

metric.py 非法篡改被拦截

策略检测到受保护评测逻辑发生变化，执行被阻断并回滚；恢复后的哈希与冻结原始值一致。

Tamper detectedYES

RollbackVERIFIED

Hash restoredYES

事实边界：这是经过校验的归档证据回放，不是实时 Agent 运行界面。页面不能执行实验、修改状态、读取任意文件或访问控制面。