File size: 6,292 Bytes
2e511b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
import csv

from legex.analysis.report import build_report


def _write(path, header, rows):
    with path.open("w", encoding="utf-8", newline="") as f:
        w = csv.writer(f)
        w.writerow(header)
        w.writerows(rows)


def test_build_report_renders_sections_and_live_numbers(tmp_path) -> None:
    iaa = tmp_path / "iaa"
    iaa.mkdir()
    analysis = tmp_path

    _write(
        iaa / "pairwise_agreement.csv",
        ["annotator_a", "annotator_b", "country", "field", "n",
         "pct_exact", "pct_tolerant", "cohen_kappa"],
        [
            # categorical (fixed-vocab) -> kappa shown
            ["primary", "a1b2c3d4e5", "de", "plaintiff_no1_ISIC1_industry_category", "20", "0.5000", "0.5000", "0.4000"],
            # monetary -> kappa computed in CSV but suppressed in the report
            ["primary", "a1b2c3d4e5", "de", "court_cost_awarded_nominal", "20", "0.6000", "0.6500", "0.3000"],
            # free text -> percent only, kappa undefined
            ["primary", "a1b2c3d4e5", "de", "legal_subject_judgement", "20", "0.1000", "0.1000", ""],
        ],
    )
    # Reference-implementation output (scripts/alt_test_reference.py); an
    # empty string marks an untestable (country, field, variant) cell.
    _write(
        iaa / "alt_test_reference_gpt-5.4-mini.csv",
        ["candidate", "country", "field",
         "winning_rate", "advantage_probability", "passes",
         "winning_rate_nontrivial", "advantage_probability_nontrivial",
         "passes_nontrivial"],
        [
            ["gpt-5.4-mini", "de", "plaintiffs_all_count",
             "1.0", "0.95", "1", "1.0", "0.93", "1"],
            ["gpt-5.4-mini", "de", "dispute_value_nominal",
             "0.3333", "0.7", "0", "", "", ""],
        ],
    )
    _write(
        analysis / "per_column.csv",
        ["model", "column", "tp", "mismatch", "missed", "hallucinated", "tn",
         "accuracy", "recall_when_filled", "precision_when_emitted",
         "hallucination_rate", "miss_rate", "wrong_when_both_filled", "f1"],
        [["gpt-5.4-mini", "plaintiffs_all_count", "8", "1", "1", "0", "10",
          "0.9000", "0.8000", "0.8889", "0.0000", "0.1000", "0.1000", "0.8400"]],
    )

    md = build_report(iaa, analysis)

    for header in (
        "# Inter-Annotator Agreement",
        "## 1. Scope",
        "## 2. Human–human agreement",
        "## 3. Alternative-annotator test",
        "## 4. Headline extraction metrics",
        "Landis–Koch",
    ):
        assert header in md, header

    assert "### 2.1 By variable" in md
    assert "0.400" in md          # ISIC (categorical) kappa is shown
    assert "0.300" not in md      # monetary kappa is suppressed, not displayed
    assert "65.0%" in md          # monetary tolerant percent agreement
    assert "±" in md              # uncertainty (±1 SE) is shown
    assert "nominal" in md and "monetary" in md  # measurement-level labels
    assert "a1b2c3d4e5" in md           # by-pair table
    assert "80.0%" in md          # recall_when_filled rendered as percent (section 4)
    assert "1/2" in md            # alt-test pass count (1 of 2 testable cells)
    assert "1/1" in md            # non-trivial pass count (untestable cell excluded)
    assert "0.82" in md           # mean advantage probability rho
    assert "0.93" in md           # non-trivial mean rho (only the testable cell)
    # provenance: numbers come from the authors' reference implementation
    assert "nitaytech/AltTest" in md
    assert "alt_test_reference.py" in md
    # per-field grids for every headline metric
    assert "### 4.3 Precision by field" in md
    assert "### 4.4 Hallucination rate by field" in md
    assert "### 4.5 F1 by field" in md
    assert "0.840" in md          # F1 rendered as 0–1, not percent
    # the alt-test formula block is always rendered, CSVs or not
    assert "### 3.1 How ω and ρ are computed" in md
    assert "rho_j = (1/n_j)" in md              # advantage probability, with its >=
    assert "(1 - epsilon) / 2  =  0.4" in md    # the threshold epsilon actually buys
    # without the decomposition CSV, §3.4 degrades to a pointer
    assert "### 3.4 Substitutable vs. better" in md
    assert "alt_test_decomposition.py" in md


def test_alt_test_decomposition_separates_substitutable_from_better(tmp_path) -> None:
    """§3.4 reports the win/tie/loss counts rho hides, and rho with ties split."""
    iaa = tmp_path / "iaa"
    iaa.mkdir()
    _write(
        iaa / "alt_test_decomposition.csv",
        ["candidate", "country", "variant", "n_comparisons",
         "refs_disagree", "llm_better", "human_better",
         "tie", "tie_same", "tie_diff", "tie_at_1", "tie_at_half", "tie_at_0",
         "rho_alttest", "rho_tiebroken"],
        [
            # ties dominate: rho looks strong, but the human wins the decisive ones
            ["gpt-5.4-mini", "de", "all", "100",
             "36", "10", "14", "76", "57", "19", "38", "26", "12", "0.86", "0.48"],
            ["gpt-5.4-mini", "de", "nontrivial", "80",
             "30", "8", "8", "64", "48", "16", "32", "22", "10", "0.90", "0.50"],
        ],
    )

    md = build_report(iaa, tmp_path)

    assert "### 3.4 Substitutable vs. better" in md
    assert "76 (76%)" in md       # tie count and share
    assert "0.86" in md           # rho as the alt-test computes it
    assert "0.48" in md           # rho once ties are split -> chance
    # the tie bucket is itemised, both by score level and by same/different answer
    assert "57 (75%)" in md       # tie_same / tie
    assert "19 (25%)" in md       # tie_diff / tie -- ties that are real disagreements
    assert "38 (50%)" in md       # tie_at_1 / tie
    assert "26 (34%)" in md       # tie_at_half / tie -- only possible if refs conflict
    assert "12 (16%)" in md       # tie_at_0 / tie
    assert "631" not in md        # refs_disagree comes from the CSV, not hard-coded
    assert "36 of 100 comparisons (36%;" in md   # reference conflict rate
    # the reading: mostly real agreement, but the human takes the decisive calls
    assert "75% of those ties are real agreement" in md
    assert "the human wins 58% (14 vs 10)" in md       # 14 / (14 + 10) decisive
    assert "0.50." in md          # non-trivial rho (ties split) survives the filter