File size: 26,306 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 | # Inter-Annotator Agreement & Alternative-Annotator Test
_Generated by `legex-analysis-report` from the CSVs in this directory — do not edit by hand; numbers always reflect the current CSVs._
Regenerate the whole chain with `scripts/reproduce_paper.sh`.
---
## 1. Scope
- Countries with ≥2 annotators: **8** — br, ch, de, ge, sg, tw, uk, us
- Annotators: **11** (primary + 10 secondary: 359f544380, 40dd94f4cc, 5c9b11ec0c, 93839740f5, 9bbad8ca9e, ab2d3cfed5, b36683f5be, c7cd8dc26a, cf5a561621, dd79b903af)
- Annotator pairs (country × pair): **24**
Countries: **br** Brazil, **ch** Switzerland, **de** Germany, **ge** Georgia, **sg** Singapore, **tw** Taiwan, **uk** United Kingdom, **us** United States.
## 2. Human–human agreement
> **Per-variable agreement.** Percent agreement is reported for every variable, using the same value-matching as the evaluation (ISO-date / number-format aware; `0` and empty treated alike). For the two **nominal ISIC** fields — the only variables with a fixed controlled vocabulary — we additionally report **Cohen's κ**. For the other variables the category space is unbounded (free text) or continuous (dates, counts, money, ratios): there κ's chance correction collapses toward percent agreement (as the number of categories → ∞, expected agreement pₑ → 0 and κ → observed agreement) and awards no partial credit for near-misses, so percent agreement is the appropriate summary. Correlation coefficients are deliberately not used — for these labels a wrong value is wrong, not partially correct. Cell detail: `kappa_audit.csv`.
_n = paired case comparisons (shared cases × annotator pairs). We report ±1 SE = √(p(1-p)/n) in percentage points, a descriptive precision marker. This is not a significance test._
### 2.1 By variable
| Variable | Type | n | % agreement | Cohen's κ |
| --- | --- | --- | --- | --- |
| `defendant_no1_ISIC1_industry_category` | nominal | 614 | 53.9% ±2.0 | 0.426 |
| `plaintiff_no1_ISIC1_industry_category` | nominal | 614 | 47.2% ±2.0 | 0.316 |
| `legal_subject_judgement` | free text | 614 | 2.4% ±0.6 | — |
| `trial_end_date` | date | 614 | 74.8% ±1.8 | — |
| `trial_start_date` | date | 614 | 70.2% ±1.8 | — |
| `plaintiffs_all_count` | count | 614 | 84.7% ±1.5 | — |
| `defendants_all_count` | count | 614 | 76.9% ±1.7 | — |
| `party_compensation_awarded_nominal` | monetary | 614 | 89.4% ±1.2 | — |
| `court_cost_awarded_nominal` | monetary | 614 | 86.0% ±1.4 | — |
| `dispute_value_nominal` | monetary | 614 | 65.3% ±1.9 | — |
| `plaintiff_loosing_share` | ratio | 614 | 72.5% ±1.8 | — |
### 2.2 By country
| Country | Pairs | n | % agreement |
| --- | --- | --- | --- |
| br | 3 | 68 | 61.1% |
| ch | 3 | 90 | 76.4% |
| de | 3 | 70 | 71.4% |
| ge | 3 | 68 | 51.7% |
| sg | 3 | 68 | 55.9% |
| tw | 3 | 70 | 71.0% |
| uk | 3 | 90 | 57.1% |
| us | 3 | 90 | 76.9% |
### 2.3 By pair
| Country | Pair | n | % agreement |
| --- | --- | --- | --- |
| br | 93839740f5 – ab2d3cfed5 | 19 | 65.5% |
| br | 93839740f5 – primary | 19 | 56.9% |
| br | ab2d3cfed5 – primary | 30 | 60.9% |
| ch | 5c9b11ec0c – ab2d3cfed5 | 30 | 71.8% |
| ch | 5c9b11ec0c – primary | 30 | 73.6% |
| ch | ab2d3cfed5 – primary | 30 | 83.6% |
| de | 40dd94f4cc – ab2d3cfed5 | 20 | 73.2% |
| de | 40dd94f4cc – primary | 20 | 75.5% |
| de | ab2d3cfed5 – primary | 30 | 67.6% |
| ge | 359f544380 – dd79b903af | 19 | 55.0% |
| ge | 359f544380 – primary | 20 | 53.6% |
| ge | dd79b903af – primary | 29 | 48.3% |
| sg | c7cd8dc26a – dd79b903af | 20 | 47.7% |
| sg | c7cd8dc26a – primary | 27 | 59.6% |
| sg | dd79b903af – primary | 21 | 58.9% |
| tw | ab2d3cfed5 – cf5a561621 | 20 | 67.7% |
| tw | ab2d3cfed5 – primary | 30 | 67.9% |
| tw | cf5a561621 – primary | 20 | 79.1% |
| uk | ab2d3cfed5 – b36683f5be | 30 | 54.8% |
| uk | ab2d3cfed5 – primary | 30 | 62.7% |
| uk | b36683f5be – primary | 30 | 53.6% |
| us | 9bbad8ca9e – ab2d3cfed5 | 30 | 76.4% |
| us | 9bbad8ca9e – primary | 30 | 76.7% |
| us | ab2d3cfed5 – primary | 30 | 77.6% |
## 3. Alternative-annotator test
> **Alternative Annotator Test** (Calderon et al. 2025, [arXiv:2501.10970](https://arxiv.org/abs/2501.10970)): leave each human annotator out in turn and score, per instance, both the candidate and the excluded annotator against the remaining annotators; a one-sided test per annotator asks whether the candidate's advantage probability trails the human's by less than ε = 0.2 (the expert-annotator tolerance), under Benjamini–Yekutieli FDR control at q = 0.05. `passes` = winning rate ≥ 0.5. Requires ≥ 3 independent annotators per country; free-text fields are excluded. ρ is the advantage probability — how likely the candidate annotates as well as or better than a randomly chosen human. The *non-trivial* variant drops instances every expert left empty: an empty prediction ties those for free, so the gap between the two columns shows how much of a pass rests on empty cells. Untestable cells (too few non-empty judgements) are excluded from the denominators.
>
> These numbers come from the **authors' reference implementation** ([github.com/nitaytech/AltTest](https://github.com/nitaytech/AltTest)) executed on the LEGEX data via `scripts/alt_test_reference.py`. See the README section "Alternative Annotator Test (AAT)" for how to run it. `legex-iaa` does not produce these CSVs.
### 3.1 How ω and ρ are computed
_Definitions from Calderon et al. (2025), as implemented in the authors' `alt_test`.
`values_agree` is the tolerant comparator LEGEX passes in as the scoring function._
```text
One jurisdiction, one candidate f, human annotators H = {1..M}, and instances i
(here: one (judgment, variable) cell over the 10 structured fields).
LEAVE-ONE-OUT. For human j and every instance i the other humans also labelled,
let A(i,-j) be the remaining humans' labels and
s_f(i,j) = score( f(i), A(i,-j) ) candidate vs. the other humans
s_h(i,j) = score( h_j(i), A(i,-j) ) held-out human vs. the same humans
score(p, A) = (1/|A|) * SUM_{a in A} values_agree(p, a) in {0, 1/2, 1}
ADVANTAGE PROBABILITY of the candidate against human j. Note the ">=", which
credits every tie to the candidate:
rho_j = (1/n_j) * SUM_i 1[ s_f(i,j) >= s_h(i,j) ]
PER-ANNOTATOR TEST. With d_i = 1[s_f < s_h] - 1[s_f >= s_h], so E[d] = 1 - 2*rho_j:
H0: E[d] >= epsilon vs. H1: E[d] < epsilon
one-sided paired t-test (Wilcoxon signed-rank when n_j < 30), the M p-values
corrected by Benjamini-Yekutieli at q = 0.05. Rejecting H0 is therefore
equivalent to rho_j being significantly greater than
(1 - epsilon) / 2 = 0.4 for the expert tolerance epsilon = 0.2
VERDICT.
omega = |{ j : H0_j rejected }| / M winning rate
rho = (1/M) * SUM_j rho_j advantage probability
"f may substitute a human annotator" <=> omega >= 0.5
```
Two consequences worth keeping in view when reading §3.2. First, ρ is a
**≥**-comparison, so a candidate that merely matches the held-out human on an
instance is scored as winning it. Second, with ε = 0.2 the hypothesis test
clears at ρ_j > 0.4, not at 0.5. Both are deliberate: the alt-test asks whether
a candidate can *substitute* a human annotator, not whether it is *better* than
one. §3.4 separates the two.
### 3.2 Headline: pooled per jurisdiction (paper numbers)
_One alt-test per jurisdiction; instance = (judgment, variable) cell over the 10 structured fields, so each annotator contributes ~190+ effective instances and the paired t-test applies without the paper's n<30 caveat._
| Candidate | Country | ω | ρ | ω (non-triv) | ρ (non-triv) |
| --- | --- | --- | --- | --- | --- |
| gpt-5.4-mini | br | 1.00 | 0.90 | 1.00 | 0.94 |
| gpt-5.4-mini | ch | 1.00 | 0.89 | 1.00 | 0.97 |
| gpt-5.4-mini | de | 0.33 | 0.80 | 0.33 | 0.78 |
| gpt-5.4-mini | ge | 1.00 | 0.82 | 1.00 | 0.84 |
| gpt-5.4-mini | sg | 1.00 | 0.86 | 1.00 | 0.87 |
| gpt-5.4-mini | tw | 1.00 | 0.86 | 1.00 | 0.87 |
| gpt-5.4-mini | uk | 1.00 | 0.81 | 1.00 | 0.80 |
| gpt-5.4-mini | us | 0.00 | 0.78 | 0.00 | 0.74 |
| gemini/gemini-3.1-flash-lite | br | 1.00 | 0.93 | 1.00 | 0.92 |
| gemini/gemini-3.1-flash-lite | ch | 1.00 | 0.90 | 1.00 | 0.93 |
| gemini/gemini-3.1-flash-lite | de | 1.00 | 0.84 | 0.33 | 0.80 |
| gemini/gemini-3.1-flash-lite | ge | 1.00 | 0.86 | 1.00 | 0.87 |
| gemini/gemini-3.1-flash-lite | sg | 1.00 | 0.88 | 1.00 | 0.90 |
| gemini/gemini-3.1-flash-lite | tw | 1.00 | 0.86 | 1.00 | 0.86 |
| gemini/gemini-3.1-flash-lite | uk | 1.00 | 0.82 | 1.00 | 0.79 |
| gemini/gemini-3.1-flash-lite | us | 0.00 | 0.75 | 0.00 | 0.69 |
| harvey | br | 1.00 | 0.91 | 1.00 | 0.87 |
| harvey | ch | 1.00 | 0.93 | 1.00 | 0.93 |
| harvey | de | 0.67 | 0.83 | 0.67 | 0.77 |
| harvey | ge | 1.00 | 0.84 | 1.00 | 0.84 |
| harvey | sg | 1.00 | 0.92 | 1.00 | 0.91 |
| harvey | tw | 1.00 | 0.85 | 1.00 | 0.82 |
| harvey | uk | 1.00 | 0.92 | 1.00 | 0.92 |
| harvey | us | 0.33 | 0.78 | 0.00 | 0.74 |
| legora-1 | br | 1.00 | 0.91 | 1.00 | 0.87 |
| legora-1 | ch | 1.00 | 0.97 | 1.00 | 0.97 |
| legora-1 | de | 0.00 | 0.70 | 0.00 | 0.57 |
| legora-1 | sg | 1.00 | 0.85 | 1.00 | 0.82 |
| legora-1 | tw | 1.00 | 0.88 | 1.00 | 0.85 |
| legora-1 | uk | 1.00 | 0.90 | 1.00 | 0.87 |
| legora-1 | us | 0.00 | 0.70 | 0.00 | 0.62 |
| legora-2 | br | 1.00 | 0.92 | 1.00 | 0.87 |
| legora-2 | ch | 1.00 | 0.96 | 1.00 | 0.95 |
| legora-2 | de | 0.00 | 0.69 | 0.00 | 0.55 |
| legora-2 | sg | 1.00 | 0.86 | 1.00 | 0.84 |
| legora-2 | tw | 1.00 | 0.88 | 1.00 | 0.85 |
| legora-2 | uk | 1.00 | 0.91 | 1.00 | 0.88 |
| legora-2 | us | 0.00 | 0.71 | 0.00 | 0.64 |
### 3.3 Per-field diagnostic (cells passed / testable)
| Candidate | ρ̄ (all) | Pass (all) | ρ̄ (non-triv) | Pass (non-triv) | br (all · non-triv) | ch (all · non-triv) | de (all · non-triv) | ge (all · non-triv) | sg (all · non-triv) | tw (all · non-triv) | uk (all · non-triv) | us (all · non-triv) |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| gemini-3.1-flash-lite | 0.86 | 47/80 | 0.85 | 40/69 | 9/10 · 6/7 | 7/10 · 7/10 | 7/10 · 4/7 | 5/10 · 6/10 | 5/10 · 6/10 | 6/10 · 6/9 | 3/10 · 2/8 | 5/10 · 3/8 |
| gpt-5.4-mini | 0.84 | 40/80 | 0.85 | 35/69 | 6/10 · 5/7 | 6/10 · 7/10 | 4/10 · 3/7 | 5/10 · 6/10 | 6/10 · 5/10 | 5/10 · 4/9 | 3/10 · 2/8 | 5/10 · 3/8 |
| harvey | 0.87 | 54/80 | 0.85 | 38/69 | 7/10 · 4/7 | 8/10 · 7/10 | 6/10 · 4/7 | 6/10 · 4/10 | 8/10 · 6/10 | 6/10 · 4/9 | 7/10 · 5/8 | 6/10 · 4/8 |
| legora-1 | 0.84 | 45/70 | 0.80 | 29/59 | 7/10 · 4/7 | 9/10 · 7/10 | 5/10 · 2/7 | – · – | 5/10 · 4/10 | 6/10 · 4/9 | 8/10 · 5/8 | 5/10 · 3/8 |
| legora-2 | 0.85 | 46/70 | 0.81 | 34/59 | 7/10 · 4/7 | 8/10 · 8/10 | 6/10 · 2/7 | – · – | 6/10 · 6/10 | 6/10 · 5/9 | 9/10 · 7/8 | 4/10 · 2/8 |
> Per-cell winning rates and advantage probabilities are in `alt_test_reference_<model>.csv`; an empty cell there means the (country, field, variant) combination was untestable.
### 3.4 Substitutable vs. better: win / tie / loss decomposition
> ρ collapses the leave-one-out comparison of §3.1 into one number, and its `≥` hands every tie to the candidate. This section keeps the same comparisons and reports the counts instead: per held-out human and instance, whether the candidate scored **better than**, **the same as**, or **worse than** that human against the two remaining humans. Restricted to judgments all three experts labelled, so every comparison has exactly two references and a score is 0, ½ or 1. `ρ (alt-test)` is the reproduced advantage probability (ties → candidate); `ρ (ties split)` counts a tie as half a win for each side. Source: `alt_test_decomposition.csv` from `scripts/alt_test_decomposition.py`, which reproduces the reference ρ of §3.2 to within 0.02.
| Candidate | Country | n | Candidate better | Tie | Human better | ρ (alt-test) | ρ (ties split) |
| --- | --- | --- | --- | --- | --- | --- | --- |
| gemini-3.1-flash-lite | br | 570 | 74 (13%) | 463 (81%) | 33 (6%) | 0.94 | 0.54 |
| gemini-3.1-flash-lite | ch | 900 | 53 (6%) | 758 (84%) | 89 (10%) | 0.90 | 0.48 |
| gemini-3.1-flash-lite | de | 600 | 37 (6%) | 470 (78%) | 93 (16%) | 0.84 | 0.45 |
| gemini-3.1-flash-lite | ge | 570 | 65 (11%) | 427 (75%) | 78 (14%) | 0.86 | 0.49 |
| gemini-3.1-flash-lite | sg | 600 | 68 (11%) | 465 (78%) | 67 (11%) | 0.89 | 0.50 |
| gemini-3.1-flash-lite | tw | 600 | 38 (6%) | 488 (81%) | 74 (12%) | 0.88 | 0.47 |
| gemini-3.1-flash-lite | uk | 900 | 87 (10%) | 650 (72%) | 163 (18%) | 0.82 | 0.46 |
| gemini-3.1-flash-lite | us | 900 | 59 (7%) | 617 (69%) | 224 (25%) | 0.75 | 0.41 |
| gemini-3.1-flash-lite | **all three** | 5640 | 481 (9%) | 4338 (77%) | 821 (15%) | 0.85 | 0.47 |
| gpt-5.4-mini | br | 570 | 72 (13%) | 445 (78%) | 53 (9%) | 0.91 | 0.52 |
| gpt-5.4-mini | ch | 900 | 46 (5%) | 759 (84%) | 95 (11%) | 0.89 | 0.47 |
| gpt-5.4-mini | de | 600 | 34 (6%) | 452 (75%) | 114 (19%) | 0.81 | 0.43 |
| gpt-5.4-mini | ge | 570 | 75 (13%) | 391 (69%) | 104 (18%) | 0.82 | 0.47 |
| gpt-5.4-mini | sg | 600 | 50 (8%) | 467 (78%) | 83 (14%) | 0.86 | 0.47 |
| gpt-5.4-mini | tw | 600 | 34 (6%) | 483 (80%) | 83 (14%) | 0.86 | 0.46 |
| gpt-5.4-mini | uk | 900 | 83 (9%) | 648 (72%) | 169 (19%) | 0.81 | 0.45 |
| gpt-5.4-mini | us | 900 | 60 (7%) | 642 (71%) | 198 (22%) | 0.78 | 0.42 |
| gpt-5.4-mini | **all three** | 5640 | 454 (8%) | 4287 (76%) | 899 (16%) | 0.84 | 0.46 |
| harvey | br | 570 | 72 (13%) | 451 (79%) | 47 (8%) | 0.92 | 0.52 |
| harvey | ch | 900 | 59 (7%) | 780 (87%) | 61 (7%) | 0.93 | 0.50 |
| harvey | de | 600 | 46 (8%) | 455 (76%) | 99 (16%) | 0.83 | 0.46 |
| harvey | ge | 570 | 79 (14%) | 399 (70%) | 92 (16%) | 0.84 | 0.49 |
| harvey | sg | 600 | 78 (13%) | 473 (79%) | 49 (8%) | 0.92 | 0.52 |
| harvey | tw | 600 | 55 (9%) | 466 (78%) | 79 (13%) | 0.87 | 0.48 |
| harvey | uk | 900 | 139 (15%) | 693 (77%) | 68 (8%) | 0.92 | 0.54 |
| harvey | us | 900 | 66 (7%) | 639 (71%) | 195 (22%) | 0.78 | 0.43 |
| harvey | **all three** | 5640 | 594 (11%) | 4356 (77%) | 690 (12%) | 0.88 | 0.49 |
| legora-1 | br | 540 | 68 (13%) | 434 (80%) | 38 (7%) | 0.93 | 0.53 |
| legora-1 | ch | 900 | 59 (7%) | 818 (91%) | 23 (3%) | 0.97 | 0.52 |
| legora-1 | de | 570 | 49 (9%) | 352 (62%) | 169 (30%) | 0.70 | 0.39 |
| legora-1 | sg | 600 | 74 (12%) | 432 (72%) | 94 (16%) | 0.84 | 0.48 |
| legora-1 | tw | 600 | 57 (10%) | 480 (80%) | 63 (10%) | 0.90 | 0.49 |
| legora-1 | uk | 900 | 128 (14%) | 679 (75%) | 93 (10%) | 0.90 | 0.52 |
| legora-1 | us | 900 | 71 (8%) | 560 (62%) | 269 (30%) | 0.70 | 0.39 |
| legora-1 | **all three** | 5010 | 506 (10%) | 3755 (75%) | 749 (15%) | 0.85 | 0.48 |
| legora-2 | br | 510 | 65 (13%) | 408 (80%) | 37 (7%) | 0.93 | 0.53 |
| legora-2 | ch | 900 | 62 (7%) | 804 (89%) | 34 (4%) | 0.96 | 0.52 |
| legora-2 | de | 600 | 50 (8%) | 362 (60%) | 188 (31%) | 0.69 | 0.39 |
| legora-2 | sg | 600 | 72 (12%) | 440 (73%) | 88 (15%) | 0.85 | 0.49 |
| legora-2 | tw | 600 | 56 (9%) | 484 (81%) | 60 (10%) | 0.90 | 0.50 |
| legora-2 | uk | 900 | 131 (15%) | 686 (76%) | 83 (9%) | 0.91 | 0.53 |
| legora-2 | us | 900 | 72 (8%) | 570 (63%) | 258 (29%) | 0.71 | 0.40 |
| legora-2 | **all three** | 5010 | 508 (10%) | 3754 (75%) | 748 (15%) | 0.85 | 0.48 |
**What is in the “Tie” bucket.** A tie only means *same score against the same two references*, so it merges several different situations. Split by score level, and independently by whether the candidate actually produced the held-out expert's answer:
| Candidate | Ties | Same answer as expert | Different answer, equal score | at 1 (all agree) | at ½ (experts conflict) | at 0 (both differ) |
| --- | --- | --- | --- | --- | --- | --- |
| gemini-3.1-flash-lite | 4338 | 3757 (87%) | 581 (13%) | 2930 (68%) | 901 (21%) | 507 (12%) |
| gpt-5.4-mini | 4287 | 3654 (85%) | 633 (15%) | 2888 (67%) | 855 (20%) | 544 (13%) |
| harvey | 4356 | 3926 (90%) | 430 (10%) | 2986 (69%) | 972 (22%) | 398 (9%) |
| legora-1 | 3755 | 3437 (92%) | 318 (8%) | 2639 (70%) | 796 (21%) | 320 (9%) |
| legora-2 | 3754 | 3446 (92%) | 308 (8%) | 2665 (71%) | 770 (21%) | 319 (8%) |
A tie at ½ is only possible when the two reference experts contradict each other — that caps every achievable score at ½, for the candidate and the held-out expert alike. The two references disagree in 1543 of 5640 comparisons (27%; a property of the human labels, identical for every candidate). Note that reference disagreement is *not* the same thing as a tie: ties also arise, and in fact more often, where the two references agree and the candidate simply matches them.
**Reading.**
- **gemini-3.1-flash-lite** — 77% of the 5640 comparisons are ties. 87% of those ties are real agreement (candidate gave the held-out expert's answer); the other 13% are comparisons where candidate and expert gave *different* answers that happened to score the same, and ρ credits every one of them to the candidate. On the 1302 comparisons that actually discriminate, the human wins 63% (821 vs 481). Dropping instances whose reference is empty throughout leaves ρ (ties split) at 0.46.
- **gpt-5.4-mini** — 76% of the 5640 comparisons are ties. 85% of those ties are real agreement (candidate gave the held-out expert's answer); the other 15% are comparisons where candidate and expert gave *different* answers that happened to score the same, and ρ credits every one of them to the candidate. On the 1353 comparisons that actually discriminate, the human wins 66% (899 vs 454). Dropping instances whose reference is empty throughout leaves ρ (ties split) at 0.47.
- **harvey** — 77% of the 5640 comparisons are ties. 90% of those ties are real agreement (candidate gave the held-out expert's answer); the other 10% are comparisons where candidate and expert gave *different* answers that happened to score the same, and ρ credits every one of them to the candidate. On the 1284 comparisons that actually discriminate, the human wins 54% (690 vs 594). Dropping instances whose reference is empty throughout leaves ρ (ties split) at 0.48.
- **legora-1** — 75% of the 5010 comparisons are ties. 92% of those ties are real agreement (candidate gave the held-out expert's answer); the other 8% are comparisons where candidate and expert gave *different* answers that happened to score the same, and ρ credits every one of them to the candidate. On the 1255 comparisons that actually discriminate, the human wins 60% (749 vs 506). Dropping instances whose reference is empty throughout leaves ρ (ties split) at 0.45.
- **legora-2** — 75% of the 5010 comparisons are ties. 92% of those ties are real agreement (candidate gave the held-out expert's answer); the other 8% are comparisons where candidate and expert gave *different* answers that happened to score the same, and ρ credits every one of them to the candidate. On the 1256 comparisons that actually discriminate, the human wins 60% (748 vs 508). Dropping instances whose reference is empty throughout leaves ρ (ties split) at 0.45.
This is what the alt-test is and is not evidence for. The tie rate is high and mostly genuine, so on this data the models are largely *indistinguishable* from an additional expert — which is exactly the substitutability claim ω and ρ are designed to support, and §3.2 supports it. It is not evidence of superiority: once ties stop counting as wins, ρ sits at chance, and on the comparisons that separate the two the human expert is still ahead. "Can this model replace a human annotator?" and "is this model better than a human annotator?" are different questions, and only the first one is being tested.
## 4. Headline extraction metrics
> Recall = TP / (TP + Mismatch + Missed) over cells the expert filled; precision = TP / (TP + Mismatch + Hallucinated); hallucination = invented values on empty-gold cells. Buckets from `legex/evaluation.py`; per-cell source `../per_column.csv`. n = evaluated label cells; recall and precision carry ±1 SE over their gold-filled / emitted denominators. The per-field grids (§4.2–4.5) break each metric out by variable across models; their `n` is the metric's model-independent denominator (gold-filled for recall, gold-empty for hallucination) and is omitted for precision (emitted; model-dependent) and F1 (composite).
### 4.1 Overall (all countries, summed across fields)
| Model | n | Accuracy | Recall (filled) | Precision | Hallu. rate | F1 |
| --- | --- | --- | --- | --- | --- | --- |
| gemini-3.1-flash-lite | 17017 | 61.5% | 57.8% ±0.5 | 57.4% ±0.5 | 30.3% | 0.576 |
| gpt-5.4-mini | 17006 | 58.5% | 58.7% ±0.5 | 54.2% ±0.4 | 41.9% | 0.563 |
| harvey | 15158 | 61.7% | 51.5% ±0.5 | 64.9% ±0.5 | 16.3% | 0.574 |
| harvey-2 | 15323 | 61.1% | 48.3% ±0.5 | 68.6% ±0.5 | 11.4% | 0.567 |
| legora-1 | 13827 | 65.0% | 52.5% ±0.5 | 73.6% ±0.5 | 8.7% | 0.613 |
| legora-2 | 13981 | 65.0% | 52.7% ±0.5 | 73.3% ±0.5 | 9.1% | 0.613 |
### 4.2 Recall (filled) by field (all countries)
| Field | n | gemini-3.1-flash-lite | gpt-5.4-mini | harvey | harvey-2 | legora-1 | legora-2 |
| --- | --- | --- | --- | --- | --- | --- | --- |
| `court_cost_awarded_nominal` | 501 | 63.9% | 65.5% | 57.9% | 57.0% | 70.1% | 70.4% |
| `defendant_no1_ISIC1_industry_category` | 1075 | 58.0% | 64.7% | 49.2% | 33.7% | 55.5% | 50.7% |
| `defendants_all_count` | 1505 | 76.9% | 73.8% | 71.6% | 71.8% | 66.6% | 65.8% |
| `dispute_value_nominal` | 708 | 56.4% | 57.1% | 47.0% | 33.1% | 34.7% | 39.7% |
| `legal_subject_judgement` | 1541 | 3.6% | 4.9% | 3.9% | 0.7% | 1.8% | 2.2% |
| `party_compensation_awarded_nominal` | 364 | 71.4% | 67.3% | 66.9% | 63.1% | 79.1% | 79.7% |
| `plaintiff_loosing_share` | 1323 | 71.3% | 69.9% | 71.4% | 69.7% | 69.3% | 70.8% |
| `plaintiff_no1_ISIC1_industry_category` | 865 | 54.1% | 53.9% | 41.8% | 31.3% | 36.8% | 36.9% |
| `plaintiffs_all_count` | 1532 | 88.8% | 85.4% | 74.9% | 77.6% | 70.5% | 72.4% |
| `trial_end_date` | 1487 | 67.6% | 66.9% | 65.0% | 65.0% | 77.8% | 77.0% |
| `trial_start_date` | 916 | 26.2% | 41.6% | 13.3% | 15.1% | 27.3% | 27.2% |
### 4.3 Precision by field (all countries)
| Field | gemini-3.1-flash-lite | gpt-5.4-mini | harvey | harvey-2 | legora-1 | legora-2 |
| --- | --- | --- | --- | --- | --- | --- |
| `court_cost_awarded_nominal` | 82.9% | 72.9% | 85.2% | 83.7% | 94.7% | 94.8% |
| `defendant_no1_ISIC1_industry_category` | 50.6% | 52.1% | 55.8% | 51.1% | 70.6% | 65.1% |
| `defendants_all_count` | 76.6% | 80.3% | 85.0% | 84.6% | 85.7% | 88.5% |
| `dispute_value_nominal` | 45.6% | 42.3% | 45.5% | 46.7% | 63.0% | 63.2% |
| `legal_subject_judgement` | 3.6% | 4.9% | 4.8% | 1.3% | 3.0% | 3.7% |
| `party_compensation_awarded_nominal` | 74.1% | 55.9% | 77.2% | 75.9% | 80.2% | 79.7% |
| `plaintiff_loosing_share` | 74.1% | 68.6% | 76.7% | 85.2% | 90.4% | 85.6% |
| `plaintiff_no1_ISIC1_industry_category` | 35.5% | 35.2% | 49.2% | 48.2% | 54.0% | 55.8% |
| `plaintiffs_all_count` | 89.1% | 89.8% | 88.6% | 89.6% | 94.5% | 93.4% |
| `trial_end_date` | 68.3% | 71.1% | 91.8% | 85.9% | 87.8% | 88.2% |
| `trial_start_date` | 58.8% | 32.4% | 57.1% | 67.4% | 78.7% | 80.7% |
### 4.4 Hallucination rate by field (all countries)
| Field | n | gemini-3.1-flash-lite | gpt-5.4-mini | harvey | harvey-2 | legora-1 | legora-2 |
| --- | --- | --- | --- | --- | --- | --- | --- |
| `court_cost_awarded_nominal` | 1046 | 4.1% | 8.3% | 1.9% | 0.8% | 0.9% | 0.7% |
| `defendant_no1_ISIC1_industry_category` | 472 | 65.9% | 74.7% | 37.3% | 27.4% | 22.4% | 28.5% |
| `defendants_all_count` | 42 | 90.5% | 56.1% | 48.6% | 62.2% | 37.1% | 18.4% |
| `dispute_value_nominal` | 839 | 44.0% | 53.8% | 32.9% | 20.4% | 12.3% | 12.9% |
| `legal_subject_judgement` | 6 | 83.3% | 100.0% | 83.3% | 0.0% | 16.7% | 16.7% |
| `party_compensation_awarded_nominal` | 1183 | 6.6% | 13.6% | 4.7% | 4.9% | 4.2% | 4.5% |
| `plaintiff_loosing_share` | 224 | 36.6% | 55.2% | 33.3% | 15.9% | 11.4% | 16.3% |
| `plaintiff_no1_ISIC1_industry_category` | 682 | 79.5% | 80.5% | 25.4% | 18.8% | 16.9% | 14.8% |
| `plaintiffs_all_count` | 15 | 100.0% | 71.4% | 83.3% | 66.7% | 28.6% | 35.7% |
| `trial_end_date` | 60 | 50.0% | 44.1% | 12.3% | 14.0% | 13.5% | 13.2% |
| `trial_start_date` | 631 | 9.7% | 61.3% | 4.4% | 3.3% | 3.6% | 3.6% |
### 4.5 F1 by field (all countries)
| Field | gemini-3.1-flash-lite | gpt-5.4-mini | harvey | harvey-2 | legora-1 | legora-2 |
| --- | --- | --- | --- | --- | --- | --- |
| `court_cost_awarded_nominal` | 0.722 | 0.690 | 0.689 | 0.679 | 0.806 | 0.808 |
| `defendant_no1_ISIC1_industry_category` | 0.540 | 0.578 | 0.523 | 0.406 | 0.621 | 0.570 |
| `defendants_all_count` | 0.768 | 0.769 | 0.777 | 0.777 | 0.750 | 0.755 |
| `dispute_value_nominal` | 0.504 | 0.486 | 0.463 | 0.387 | 0.447 | 0.488 |
| `legal_subject_judgement` | 0.036 | 0.049 | 0.043 | 0.009 | 0.023 | 0.028 |
| `party_compensation_awarded_nominal` | 0.727 | 0.611 | 0.717 | 0.689 | 0.797 | 0.797 |
| `plaintiff_loosing_share` | 0.727 | 0.692 | 0.740 | 0.767 | 0.784 | 0.775 |
| `plaintiff_no1_ISIC1_industry_category` | 0.429 | 0.426 | 0.452 | 0.379 | 0.438 | 0.444 |
| `plaintiffs_all_count` | 0.889 | 0.875 | 0.812 | 0.832 | 0.807 | 0.816 |
| `trial_end_date` | 0.679 | 0.690 | 0.761 | 0.740 | 0.825 | 0.822 |
| `trial_start_date` | 0.362 | 0.364 | 0.216 | 0.247 | 0.405 | 0.406 |
## Appendix
### Landis–Koch (1977) κ scale (for interpreting the ISIC κ)
| κ | Strength |
| --- | --- |
| < 0.00 | Poor (worse than chance) |
| 0.00 – 0.20 | Slight |
| 0.21 – 0.40 | Fair |
| 0.41 – 0.60 | Moderate |
| 0.61 – 0.80 | Substantial |
| 0.81 – 1.00 | Almost perfect |
### Provenance
- Pairwise agreement & kappa: `legex/analysis/iaa.py` (`pairwise_agreement`, `write_kappa_audit_csv`).
- Alt-test: authors' reference implementation ([github.com/nitaytech/AltTest](https://github.com/nitaytech/AltTest)) run via `scripts/alt_test_reference.py` → `alt_test_reference_*.csv` (see README).
- Alt-test win/tie/loss decomposition: `scripts/alt_test_decomposition.py` → `alt_test_decomposition.csv`.
- Tolerant comparator: `legex/evaluation/comparison.py` (`values_agree`, `normalise`).
- Headline buckets: `legex/analysis/aggregate.py` over `legex/evaluation.score_country`.
- This report: `legex/analysis/report.py`.
|