File size: 4,848 Bytes
cb288cf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
{
  "utc": "2026-10-21T01:19:09.758641+00:00",
  "timing": "specified during inference, before full cross-arm analysis or scientific content scoring",
  "primary_arms": [
    "qwen_base",
    "qwen_final",
    "gemma_base",
    "gemma_final"
  ],
  "contrasts": [
    {
      "name": "qwen_finetuning",
      "left": "qwen_base",
      "right": "qwen_final"
    },
    {
      "name": "gemma_finetuning",
      "left": "gemma_base",
      "right": "gemma_final"
    },
    {
      "name": "base_cross_model",
      "left": "qwen_base",
      "right": "gemma_base"
    },
    {
      "name": "final_cross_model",
      "left": "qwen_final",
      "right": "gemma_final"
    }
  ],
  "secondary_generation_contrasts": [
    {
      "name": "qwen_domain_adaptation",
      "left": "qwen_base",
      "right": "qwen_domain"
    },
    {
      "name": "gemma_domain_adaptation",
      "left": "gemma_base",
      "right": "gemma_domain"
    },
    {
      "name": "qwen_semantic_stage",
      "left": "qwen_domain",
      "right": "qwen_final"
    },
    {
      "name": "gemma_semantic_stage",
      "left": "gemma_domain",
      "right": "gemma_final"
    }
  ],
  "difference": "right minus left",
  "generation_metrics": {
    "json": [
      "json_valid",
      "reference_top_level_keys_present",
      "exact_json_reference_match"
    ],
    "tool": [
      "tool_schema_valid",
      "tool_names_match_reference",
      "tool_arguments_match_reference"
    ],
    "text": [
      "nonempty"
    ]
  },
  "source_metrics": {
    "source_card": [
      "json_valid",
      "identifiers_match",
      "schema_fields_present",
      "all_nonempty_spans_supported",
      "numeric_literals_exact",
      "mechanical_grounding_pass"
    ],
    "quote_locator": [
      "json_valid",
      "identifiers_match",
      "quote_match",
      "interval_matches_source",
      "exact_reference_position",
      "mechanical_grounding_pass"
    ]
  },
  "workflow_primary_endpoints": {
    "delivery_complete": "原生终态同时具有非占位实验计划与待审报告,并覆盖材料/工艺/性能能力/安全/综合评价/实验规划角色;人工批准不自动执行。",
    "scientifically_usable_delivery": "delivery_complete且有两个实质不同方向、原文经典对照、可检验假设、对应目标的实验测量与样品身份绑定;未识别实质事实/数值/条件/引用错误,且没有阻止执行的未解决身份或方法缺口。正确声明输入不足、未计算、未实验可接受。",
    "scientific_content_score": "来源蕴含、身份数值条件、假设与对照、测量设计、执行与能力边界5维,各0/1/2;0为错误或完全缺失,1为部分或未解决,2为明确满足。对最终可交付内容评分,未交付记0并保留原稿诊断。"
  },
  "workflow_review": {
    "assessor": "本助手的自动/人工阅读式审查,非领域专家",
    "blinding": "打乱案例和臂顺序并生成无模型/协调标签的原答与最终状态文本;运行轨迹仅在内容评分锁定后核验。执行者已知研究设计,不能声称完全双盲。",
    "expert_review_completed": false,
    "expert_decision": null,
    "rule": "每项实质问题给出输入来源和模型原文定位;格式失败不自动等同科学判断全错;未审项null。"
  },
  "bootstrap": {
    "replicates": 10000,
    "seed": 20261018,
    "confidence": 0.95,
    "method": "paired cluster percentile; resample original case-group connected components for generation, DOI for source/workflow; numerator/denominator aggregated within each resample",
    "single_cluster": "report paired point difference without confidence interval"
  },
  "workflow_binary_tests": "two-sided exact McNemar on12 DOI pairs; Holm correction across four primary contrasts times two binary endpoints",
  "generation_source_tests": "paired cluster confidence intervals and effect sizes; no independent assistant-turn binomial test",
  "workflow_science_score": "five dimensions each0/1/2, sum0..10; missing final delivery scored0 per inherited rubric; partial raw-content diagnostics reviewed separately",
  "incomplete_policy": "missing/unexecuted remains null, no complete comparison until expected paired cohort and integrity records are present",
  "content_review": "blinded output IDs, original task and sources, final result and unmodified generated content except terminal EOS; metadata labels hidden; assessor knows study design, not fully double blind; lock scores before unblinding",
  "expert_review_completed": false,
  "timing_cost_policy": "record native token counts and elapsed costs; historical reused time includes decoding/scoring while fresh generation time ends before decoding/scoring. Show origin and timing scope; no high-precision speed superiority claim across these scopes."
}