[code] Reproduction bundle.

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+*.jsonl filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -3,6 +3,8 @@ __pycache__/
 *.py[codz]
 *$py.class
 *.pdf
+# ... but the shipped appendix figure is part of the release payload
+!data/analysis/figures/isic_frequencies.pdf
 
 # C extensions
 *.so

HUGGINGFACE_UPLOAD.md

@@ -0,0 +1,257 @@
+# How to publish LEGEX to Hugging Face anonymously
+
+This walkthrough takes the contents of `submission/` and publishes **three
+public Hugging Face dataset repositories** under the anonymous account
+`legexbenchmark`:
+
+1. **`legexbenchmark/goldensets`** — expert-coded gold annotations
+   (19 jurisdictions, 1,548 judgments / 1,974 rows incl. secondary
+   re-annotations).
+2. **`legexbenchmark/inference-results`** — Harvey (2 runs) / Gemini / GPT /
+   Legora (2 runs) outputs (19 jurisdictions × 6 runs = 114 files, JSONL)
+   plus the per-column prompts of the commercial runs (`prompts/`, 4 files).
+3. **`legexbenchmark/code`** — Python pipeline (scrape → process → infer →
+   clean → evaluate).
+
+> **All data files are JSONL** (one JSON object per line). Inference was CSV
+> in the first release and is now JSONL — see *Updating an existing release*
+> below.
+
+---
+
+## 0. Updating an existing release (goldensets + inference-results)
+
+The three repos already exist, and local clones live in
+`../hf-uploads/{goldensets,inference-results,code}`. The staged bundle
+(`submission/goldensets/`, `submission/inference-results/`) mirrors the two
+data repos 1:1: data + the per-repo dataset card (`README.md`).
+
+Rebuild the bundle only if the underlying data changed
+(`convert_goldenset_to_jsonl.py` needs the local-only
+`data/reannotation/annotators.json` + `ANNOTATOR_SALT` from `.env`):
+
+```bash
+uv run python submission/convert_goldenset_to_jsonl.py --data-dir data --out-dir submission/goldensets/data
+(cd submission && uv run python build_inference_jsonl.py --data-dir ../data --gold-dir goldensets/data --out-dir inference-results/data)
+```
+
+When the working trees in `../hf-uploads/` are already synced from
+`submission/` (data + README copied in, stale CSVs removed), pushing the
+update is:
+
+```bash
+# Log in once with the anonymous account's WRITE token.
+pip install --upgrade huggingface_hub && git lfs install --skip-repo
+huggingface-cli login                 # paste the hf_… write token
+huggingface-cli whoami                # must print: legexbenchmark
+
+# --- goldensets ---
+cd ../hf-uploads/goldensets
+git status && git diff --stat         # review: data/<cc>/*.jsonl + README.md
+git config user.name anonymous && git config user.email anonymous@anonymous.invalid
+git add -A
+git commit -m "Update: re-annotations, Nepal/Singapore extensions, traceability fields, refreshed card"
+git push
+cd -
+
+# --- inference-results (CSV -> JSONL) ---
+cd ../hf-uploads/inference-results
+git status                            # review: CSVs deleted, 57 JSONL added, README.md, .gitattributes
+git config user.name anonymous && git config user.email anonymous@anonymous.invalid
+git add -A
+git commit -m "Switch inference to JSONL; 19 jurisdictions; cleaning provenance; refreshed card"
+git push
+cd -
+```
+
+Then **verify usability** (the old CSV raised `ArrowInvalid`):
+
+```python
+from datasets import load_dataset
+load_dataset("legexbenchmark/goldensets", split="brazil")
+load_dataset("legexbenchmark/inference-results", "harvey", split="united_states")
+load_dataset("legexbenchmark/inference-results", "gpt", split="switzerland")
+```
+
+Also confirm on the web UI that every commit author reads **`anonymous`**, the
+dataset viewer renders both repos, and no `inference_*.csv` remain in
+`inference-results`.
+
+**`code`** is updated separately once the cleaned analysis package is staged —
+see item 3 of [CHANGES_SINCE_SUBMISSION.md](CHANGES_SINCE_SUBMISSION.md).
+
+---
+
+## 1. Create the anonymous Hugging Face account (once)
+
+1. Open a private / incognito browser window so cookies from any existing
+   HF account are isolated.
+2. Visit <https://huggingface.co/join>.
+3. Sign up with a **fresh email address** that is not tied to any of the
+   authors (a `proton.me` / `tutanota.com` mailbox works fine).
+4. Pick the username **`legexbenchmark`** so the public URLs match the
+   names used in the paper.
+5. Verify the email and log in.
+
+> **Why a fresh account?** HF attaches your account's display name, avatar
+> and email to commits in each repo's history. Reusing an existing account
+> would leak your identity even if every file in the repo is clean.
+
+## 2. Create a write-scoped access token (once)
+
+1. Visit <https://huggingface.co/settings/tokens>.
+2. Click **New token**, name it `upload`, role **Write**, click **Generate
+   token**, copy the value (it starts with `hf_`).
+
+## 3. Install the tooling and log in (once)
+
+```bash
+pip install --upgrade huggingface_hub
+git lfs install --skip-repo   # macOS: brew install git-lfs first
+
+# Paste the write token created in step 2 when prompted.
+huggingface-cli login
+```
+
+## 4. Push the three repositories (initial release)
+
+Run steps 4a → 4c **in order**. Each step is independent — if something
+breaks in 4b you can fix it without redoing 4a.
+
+### 4a. Push `legexbenchmark/goldensets`
+
+```bash
+# 1. Create the empty repo via the web UI:
+#    https://huggingface.co/new-dataset
+#    Owner: legexbenchmark | Name: goldensets | License: MIT | Public
+
+# 2. Clone, anonymise local git, copy in the staged files.
+git clone https://huggingface.co/datasets/legexbenchmark/goldensets hf-goldensets
+cd hf-goldensets
+git config user.name  "anonymous"
+git config user.email "anonymous@anonymous.invalid"
+git lfs track "*.jsonl"
+git add .gitattributes
+
+cp -R ../submission/goldensets/. .
+
+# 3. Commit and push.
+git add .
+git commit -m "Initial release"
+git push -u origin main
+cd ..
+```
+
+Open <https://huggingface.co/datasets/legexbenchmark/goldensets>. Verify:
+
+- The dataset card renders with the jurisdiction tables.
+- `data/<cc>/goldenset_<cc>.jsonl` is browsable for all 19 jurisdictions.
+- The author next to every commit is **`anonymous`**.
+
+### 4b. Push `legexbenchmark/inference-results`
+
+```bash
+# 1. Create the empty repo:
+#    https://huggingface.co/new-dataset
+#    Owner: legexbenchmark | Name: inference-results | License: MIT | Public
+
+# 2. Clone, anonymise, copy in.
+git clone https://huggingface.co/datasets/legexbenchmark/inference-results hf-inference
+cd hf-inference
+git config user.name  "anonymous"
+git config user.email "anonymous@anonymous.invalid"
+git lfs track "*.jsonl"
+git add .gitattributes
+
+cp -R ../submission/inference-results/. .
+
+# 3. Commit and push.
+git add .
+git commit -m "Initial release"
+git push -u origin main
+cd ..
+```
+
+Verify on <https://huggingface.co/datasets/legexbenchmark/inference-results>:
+
+- 19 `data/<cc>/` folders (all jurisdictions).
+- Each folder has `inference_harvey.jsonl`, `inference_gemini.jsonl`,
+  `inference_gpt.jsonl` (no `.csv`).
+
+### 4c. Push `legexbenchmark/code`
+
+```bash
+# 1. Create the empty repo:
+#    https://huggingface.co/new-dataset
+#    Owner: legexbenchmark | Name: code | License: MIT | Public
+#    (HF dataset repos accept arbitrary files — we use "dataset" type as
+#    the most permissive container for source code.)
+
+# 2. Clone, anonymise, copy in the cleaned package (see
+#    CHANGES_SINCE_SUBMISSION.md item 3 for what goes into the code repo).
+git clone https://huggingface.co/datasets/legexbenchmark/code hf-code
+cd hf-code
+git config user.name  "anonymous"
+git config user.email "anonymous@anonymous.invalid"
+
+# 3. Commit and push.
+git add .
+git commit -m "Initial release"
+git push -u origin main
+cd ..
+```
+
+Verify on <https://huggingface.co/datasets/legexbenchmark/code>:
+
+- The repo card lists the CLI entrypoints and the workflow.
+- `legex/scrapers/` contains 19 jurisdiction modules.
+
+## 5. Cross-verify the org page
+
+Open <https://huggingface.co/legexbenchmark> in a fresh tab — you should
+see three pinned datasets. Click each card and confirm:
+
+- The README renders correctly.
+- The "Files and versions" tab is browsable.
+- No identifying information appears next to commits (everything should
+  read `anonymous`).
+
+Also run one final string search per repo via the file viewer for your own
+surname, your affiliation, and your work email domain:
+
+```
+<your-surname>       # must return nothing
+<your-affiliation>   # must return nothing
+<your-email-domain>  # must return nothing
+```
+
+## 6. Share with reviewers
+
+The single URL that ties the three repos together is
+<https://huggingface.co/legexbenchmark>. Add it to the supplementary
+material of your OpenReview submission; reviewers can browse all three
+datasets from that org page without logging into HF.
+
+---
+
+## Troubleshooting
+
+**A commit shows your real name despite the local config.** Run
+`git log -1 --pretty='%an <%ae>'` inside the clone. If it still shows your
+real identity, redo `git config user.name / user.email` and then
+`git commit --amend --reset-author` before pushing.
+
+**HF complains about LFS quota.** Free accounts get 1 GB of LFS storage;
+this release uses ~30 MB so quota is not a concern.
+
+**A file exceeds 5 GB.** None of the files in this release do. If a future
+addition does, `git lfs track` the path and recommit.
+
+**You want to delete a repo.** Each repo's Settings page has "Delete this
+dataset repository" — that removes files and history.
+
+**You need to re-push after a fix.** `cd` into the local clone, edit, commit
+with `--reset-author` if you also re-ran `git config`, then `git push`.
+
+**Reviewers ask for a single zip.** HF's "Use this dataset" → "Repository
+size" → "Download repository" gives a `.tar.gz` of the whole repo.

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2026 Anonymous Authors
+Copyright (c) 2026 Engineers for Science
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

README.md

@@ -1,91 +1,183 @@
----
-license: mit
-pretty_name: "LEGEX Code: Scrapers, Inference and Evaluation Pipeline"
-tags:
-  - legal
-  - benchmark
-  - code
-  - llm-evaluation
-  - information-extraction
----
-
-# LEGEX Code, Scrapers, Inference and Evaluation Pipeline
-
-Python source for the LEGEX benchmark of civil-judgment review-table
-extraction. This repository contains:
-
-- Scrapers for 19 jurisdictions (per-court HTML / API / HuggingFace
-  pull) in [`legex/scrapers/`](legex/scrapers/).
-- Inference pipeline that calls Harvey, Gemini and OpenAI APIs against
-  a schema-constrained 14-field review table
-  ([`legex/inference.py`](legex/inference.py),
-  [`legex/harvey.py`](legex/harvey.py),
-  [`legex/models/classification.py`](legex/models/classification.py)).
-- Evaluation that compares system outputs against expert-coded gold
-  cells ([`legex/evaluation.py`](legex/evaluation.py)), aggregates across
-  jurisdictions ([`legex/analysis.py`](legex/analysis.py)), and renders
-  paper tables ([`legex/quant_results.py`](legex/quant_results.py)).
-- Conversion script [`convert_goldenset_to_jsonl.py`](convert_goldenset_to_jsonl.py)
-  — turns the source XLSX goldensets into the JSONL format used by
-  [`legexbenchmark/goldensets`](https://huggingface.co/datasets/legexbenchmark/goldensets).
-
-
-## Setup
+# LEGEX — Reproduction Bundle (release/icml2026)
+
+LEGEX is an expert-coded benchmark for civil-judgment review-table
+extraction: judgments from the highest civil courts of 19 jurisdictions,
+annotated by legal experts on a shared schema (monetary outcomes, cost
+allocation, party structure, industry classification), against which we
+evaluate two commercial review-table systems (Harvey, Legora) and two
+schema-constrained LLM pipelines (Gemini, ChatGPT). This bundle is the
+versioned code + data payload behind the paper LEGEX: An Eight-Jurisdiction
+Benchmark for Legal Review-Table Extraction (ICML 2026 Workshop on AI for
+Law). It contains an eight-jurisdiction core benchmark (≥100 judgments each,
+double-annotated on a 28–30-case overlap) and eleven preview jurisdictions.
+
+The release consists of three Hugging Face repositories, pinned to the
+`release/icml2026` branch:
+
+| Artifact | Contents | URL |
+|---|---|---|
+| `goldensets` | Expert gold labels, `data/<cc>/goldenset_<cc>.jsonl` | <https://huggingface.co/datasets/legexbenchmark/goldensets/tree/release/icml2026> |
+| `inference-results` | System predictions, `data/<cc>/inference_{harvey,harvey_2,gemini,gpt,legora_1,legora_2}.jsonl`, plus the as-run per-column prompts of the commercial runs in `prompts/` | <https://huggingface.co/datasets/legexbenchmark/inference-results/tree/release/icml2026> |
+| `code` | This bundle (package, scripts, shipped analysis outputs) | <https://huggingface.co/datasets/legexbenchmark/code/tree/release/icml2026> |
+
+## Layout of this bundle
+
+```
+submission/
+├── README.md                       this file
+├── pyproject.toml · LICENSE · .env.template
+├── legex/                          the LEGEX package, pruned to what the release uses:
+│   ├── scrapers/                   one scraper per jurisdiction
+│   ├── fulltext/                   full-text acquisition for the goldenset workbooks
+│   ├── prompts/                    versioned system prompts (paper runs use v3)
+│   ├── inference.py                legex-classify (LLM pipelines)
+│   ├── harvey.py                   legex-harvey-ingest (review-table export ingest)
+│   ├── evaluation/                 tolerant cell comparator + scoring engine
+│   └── analysis/                   aggregates, IAA, AAT rendering, report, plots
+├── scripts/                        paper-facing scripts (tables, figures, AAT adapter)
+├── tests/                          unit tests (uv run pytest -q)
+├── data/analysis/                  SHIPPED analysis outputs (see "Recompute" below)
+│   ├── per_country_per_column.csv · per_country.csv · per_column.csv
+│   ├── per_tradition.csv · per_language.csv
+│   ├── quant_results.tex           paper Table tab:overall
+│   ├── paper_tables.tex            tab:metrics-by-jurisdiction / tab:metrics-by-field
+│   ├── tables/                     diversity.tex, currency_frequencies.tex, ...
+│   ├── figures/                    isic_frequencies.pdf, legal_subject_wordcloud.png
+│   ├── iaa/                        ANALYSIS.md, pairwise_agreement.csv, kappa_audit.csv,
+│   │                               alt_test_pooled.csv, alt_test_reference_*.csv
+│   ├── hallucinations/             hallucination_review_ch.csv (hand-classified error case study)
+│   └── quality/                    by_country.csv, by_variable.csv
+├── goldensets/                     1:1 copy of legexbenchmark/goldensets (dataset card + data)
+├── inference-results/              1:1 copy of legexbenchmark/inference-results
+├── convert_goldenset_to_jsonl.py   XLSX gold workbooks -> goldenset_<cc>.jsonl
+��── build_inference_jsonl.py        cleaned working JSONL -> inference_<model>.jsonl
+├── docs/Jurisdictions.md           per-jurisdiction research catalogue
+└── HUGGINGFACE_UPLOAD.md           how this bundle maps onto the three HF repos
+```
+
+## Reproduce the paper tables and figures
+
+All commands are run **from this directory** (the bundle root). Requires
+[uv](https://docs.astral.sh/uv/) and Python ≥ 3.11.
 
 ```bash
-git clone https://huggingface.co/datasets/legexbenchmark/code legex-code
-cd legex-code
 uv sync
-cp .env.template .env
 ```
 
-Required tokens depend on which scrapers / models you run, see
-[`.env.template`](.env.template).
+**Table `tab:overall` (headline metrics).** Re-render from the shipped
+per-cell aggregate and check it is byte-identical to the shipped table:
 
-## End-to-end workflow
+```bash
+uv run legex-quant-results --input data/analysis/per_country_per_column.csv --out /tmp/quant_results.tex
+diff /tmp/quant_results.tex data/analysis/quant_results.tex
+```
+
+**Tables `tab:metrics-by-jurisdiction` and `tab:metrics-by-field`.**
 
 ```bash
-# Acquire raw judgments per jurisdiction.
-uv run legex-run
+uv run python scripts/paper_tables.py > /tmp/paper_tables.tex
+diff /tmp/paper_tables.tex data/analysis/paper_tables.tex
+```
 
-# Run inference for one system on one jurisdiction, Harvey has do be done separately as this is a commercial tool
-uv run legex-classify --country us --model gpt-5.4-mini --full_text
+**IAA / kappa / Alternative Annotator Test report** (`ANALYSIS.md`, the
+source of the paper's `tab:iaa` and AAT numbers). The report is rendered
+purely from the shipped CSVs:
 
-# Evaluate one system on one jurisdiction.
-uv run legex-evaluate --country us --system gpt
+```bash
+uv run legex-analysis-report --iaa-dir data/analysis/iaa --analysis-dir data/analysis --out /tmp/ANALYSIS.md
+diff /tmp/ANALYSIS.md data/analysis/iaa/ANALYSIS.md
+```
 
-# Aggregate across all 12 evaluated jurisdictions and 3 systems.
-uv run legex-analysis --out data/analysis
+**One-shot reproduction.** `scripts/reproduce_paper.sh` regenerates every
+artifact under `data/analysis/` — IAA CSVs, scoring aggregates, all tables,
+the diversity/frequency figures, the hallucination shares, and `ANALYSIS.md`
+— from the published bundles in `goldensets/data/` and
+`inference-results/data/`. The run is byte-stable, so a clean
+`git status data/analysis` afterwards is the verification that the shipped
+numbers reproduce:
 
-# Render the paper-headline LaTeX table.
-uv run legex-quant-results \
-    --input data/analysis/per_country_per_column.csv \
-    --out   data/analysis/quant_results.tex
+```bash
+bash scripts/reproduce_paper.sh
+git status data/analysis
 ```
 
-To evaluate against the published goldensets and inference outputs, pull
-the two data repos into the expected layout:
+## Unpublished ata
+
+The expert-annotated XLSX workbooks (`data/<cc>/Goldenset_*_final*.xlsx`)
+behind the goldensets arenot published: Microsoft Office embeds author
+metadata, and the re-annotation assignment structure could identify
+individual annotators. The gold labels are published in full as
+`goldensets/data/<cc>/goldenset_<cc>.jsonl` with primary + anonymized secondary
+annotator rows, and every analysis in this bundle runs from those published
+files, the workbooks are not needed for reproduction.
+
+## Alternative Annotator Test (AAT).
+The AAT (Calderon, Reichart & Dror,
+ACL 2025, [arXiv:2501.10970](https://arxiv.org/abs/2501.10970)) is run with
+the authors' original implementation via the adapter
+`scripts/alt_test_reference.py`, which extracts `alt_test()` from the
+upstream notebook at runtime and feeds it LEGEX data from the published
+bundles:
 
 ```bash
-huggingface-cli download legexbenchmark/goldensets         --repo-type dataset --local-dir data --include "data/*"
-huggingface-cli download legexbenchmark/inference-results  --repo-type dataset --local-dir data --include "data/*"
-# After these, data/<cc>/ contains goldenset_<cc>.jsonl + inference_*.csv
-uv run legex-analysis --out data/analysis
+git clone https://github.com/nitaytech/AltTest /tmp/AltTest
+ALTTEST_DIR=/tmp/AltTest bash scripts/reproduce_paper.sh   # or:
+uv run python scripts/alt_test_reference.py --alttest /tmp/AltTest \
+    --gold-dir goldensets/data --inference-dir inference-results/data \
+    [--countries ge,sg,tw] [--epsilon 0.2] [--out data/analysis/iaa] [--per-field]
 ```
 
-## CLI entrypoints
+Its outputs (`alt_test_pooled.csv`, `alt_test_reference_*.csv`) are shipped
+under `data/analysis/iaa/` and rendered into `ANALYSIS.md` by
+`legex-analysis-report`.
 
-| Command | Module | Purpose |
-|---|---|---|
-| `legex-run` | `legex.main:main` | Top-level scrape + filter + sample pipeline. |
-| `legex-classify` | `legex.inference:main` | Run an LLM over the goldenset and write predictions to CSV. |
-| `legex-harvey-ingest` | `legex.harvey:main` | Ingest a Harvey Vault Review export into the per-jurisdiction CSV format. |
-| `legex-evaluate` | `legex.evaluation:main` | Per-country, per-field bucket counts and recall / hallucination. |
-| `legex-analysis` | `legex.analysis:main` | Cross-jurisdiction analysis → CSV + LaTeX tables. |
-| `legex-quant-results` | `legex.quant_results:main` | Paper-headline summary from the analysis CSV. |
-| `legex-pdf` | `legex.pdf_export.cli:main` | Render per-row PDFs from a goldenset workbook. |
-| `legex-plots` | `legex.plots:main` | Plot helpers used in the paper. |
+## Scraping
+
+The scrapers that built the corpus are included (`legex/scrapers/`, one file
+per jurisdiction; see `docs/Jurisdictions.md` for the per-country catalogue).
+
+```bash
+cp .env.template .env
+uv run legex-run
+```
+
+`legex-run` executes the idempotent pipeline (scrape → filter/sample →
+fill goldenset → dist) for every registered jurisdiction and skips countries
+whose access is not set up or whose output already exists. Some sources need
+manual prerequisites (all keys go into `.env`, see
+[.env.template](.env.template)):
+
+- **Hugging Face** (`au`, `ch`, `de`, `us`): a read token as `HF_TOKEN`;
+  `de` and `us` additionally require accepting the gated-dataset terms of
+  `openlegaldata/court-decisions-germany` and
+  `free-law/Caselaw_Access_Project`.
+- **France**: register at <https://piste.gouv.fr>, subscribe to the
+  Judilibre API, set `JUDILIBRE_CLIENT_ID` / `JUDILIBRE_CLIENT_SECRET`.
+- **New Zealand**: copy the WAF cookie from a justice.govt.nz session into
+  `NZ_WAF_COOKIE`.
+- **India**: `legex-india-extract` pulls selected PDFs from the AWS Open
+  Data TAR archives (see `legex/scrapers/in_.py`).
+
+Model inference over the scraped full texts (the paper runs use prompt v3):
+
+```bash
+uv run legex-classify --model <model> --full_text --prompt_version v3
+```
+
+Tests: `uv run pytest -q`.
+
+## Citation
+
+```bibtex
+@inproceedings{legex2026,
+  title     = {{LEGEX}: An Eight-Jurisdiction Benchmark for Legal Review-Table Extraction},
+  author    = {K{\"o}nig, Adrian and others},
+  booktitle = {ICML 2026 Workshop on AI for Law (AI4Law)},
+  year      = {2026},
+  url       = {https://huggingface.co/legexbenchmark}
+}
+```
 
 ## License
 
-MIT.
+MIT. See [LICENSE](LICENSE).

build_harvey_prompts_jsonl.py

@@ -0,0 +1,86 @@
+"""Build the published Harvey prompt-metadata JSONL from the VAULT exports.
+
+Each Harvey VAULT_REVIEW export carries the as-configured per-column prompt in
+its question headers ("1. 1. Case ID (<prompt>)"). This script extracts the
+value-column header of each of the 12 question blocks (same block layout as
+``legex.harvey``) and writes one record per (run, field):
+
+    {"model": "harvey", "inference_date": "2026-05-18", "field": ..., "prompt": ...}
+
+``prompts_harvey.jsonl`` holds both dates of the paper run (2026-05-18 and
+2026-06-30 — the 30 June re-created tables carry platform-rephrased prompts);
+``prompts_harvey_2.jsonl`` holds the 2026-08-05 transparency run.
+
+    python build_harvey_prompts_jsonl.py --raw-dir ../data/raw \\
+        --out-dir inference-results/prompts
+"""
+import argparse
+import json
+import re
+from pathlib import Path
+
+import openpyxl
+
+from legex.harvey import HARVEY_FIELDS_ORDER
+
+# run -> [(inference_date, export file)]
+RUNS: dict[str, tuple[tuple[str, str], ...]] = {
+    "harvey": (
+        ("2026-05-18", "harvey_2026-05-18.xlsx"),
+        ("2026-06-30", "harvey_2026-06-30.xlsx"),
+    ),
+    "harvey-2": (("2026-08-05", "harvey_2026-08-05.xlsx"),),
+}
+_FIRST_ANSWER_COL = 3  # after Name, Folder, Document Classification
+_TITLE_PREFIX_RE = re.compile(r"^\s*(?:\d+\.\s*)+")
+
+
+def _parse_header(header: str) -> str:
+    """The prompt is the parenthetical after the column title; parens may nest."""
+    start = header.find("(")
+    if start == -1 or not header.rstrip().endswith(")"):
+        raise ValueError(f"header without a prompt parenthetical: {header[:80]!r}")
+    prompt = header[start + 1 : header.rindex(")")]
+    return " ".join(prompt.split())
+
+
+def read_prompts(xlsx: Path) -> dict[str, str]:
+    """Return ``{field: prompt}`` from the value-column headers of one export."""
+    wb = openpyxl.load_workbook(xlsx, read_only=True)
+    ws = wb["Sheet1"]f
+    header = next(ws.iter_rows(min_row=1, max_row=1, values_only=True))
+    wb.close()
+    width = (len(header) - _FIRST_ANSWER_COL) // len(HARVEY_FIELDS_ORDER)
+    if width < 1:
+        raise ValueError(f"unexpected Harvey sheet width in {xlsx.name}: {len(header)} columns")
+    out: dict[str, str] = {}
+    for j, field in enumerate(HARVEY_FIELDS_ORDER):
+        cell = header[_FIRST_ANSWER_COL + j * width]
+        out[field] = _parse_header(str(cell))
+    return out
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
+    parser.add_argument("--raw-dir", type=Path, default=Path("../data/raw"),
+                        help="Directory holding the harvey_<date>.xlsx exports.")
+    parser.add_argument("--out-dir", type=Path, default=Path("inference-results/prompts"))
+    args = parser.parse_args(argv)
+
+    args.out_dir.mkdir(parents=True, exist_ok=True)
+    for run, exports in RUNS.items():
+        dst = args.out_dir / f"prompts_{run.replace('-', '_')}.jsonl"
+        n = 0
+        with dst.open("w", encoding="utf-8") as f:
+            for inference_date, filename in exports:
+                for field, prompt in read_prompts(args.raw_dir / filename).items():
+                    record = {"model": run, "inference_date": inference_date,
+                              "field": field, "prompt": prompt}
+                    f.write(json.dumps(record, ensure_ascii=False) + "\n")
+                    n += 1
+        print(f"wrote {n} prompt record(s) to {dst}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

build_inference_jsonl.py

@@ -0,0 +1,56 @@
+"""Build the published inference-results bundle from the cleaned working JSONL.
+
+For every annotated jurisdiction (one with a ``goldenset_<cc>.jsonl`` under
+``--gold-dir/<cc>/``) copy each model's cleaned inference file
+``<data-dir>/<cc>/Goldenset_*_v3_full_text_<slug>.jsonl`` to
+``--out-dir/<cc>/inference_<model>.jsonl``. The records are already clean,
+single-typed JSONL with ``comment`` / ``original_input`` provenance (see
+``legex.evaluation.cleaning``), so this is a faithful copy — no transformation,
+no CSV type-inference hazard.
+
+    python build_inference_jsonl.py --data-dir ../data \\
+        --gold-dir goldensets/data --out-dir inference-results/data
+"""
+import argparse
+import re
+from pathlib import Path
+
+# working model slug -> published short name
+SLUG_TO_MODEL = {
+    "harvey": "harvey",
+    "harvey-2": "harvey_2",
+    "gpt-5.4-mini": "gpt",
+    "gemini_gemini-3.1-flash-lite": "gemini",
+    "legora-1": "legora_1",
+    "legora-2": "legora_2",
+}
+_SLUG_RE = re.compile(r"_v3_full_text_(.+)\.jsonl$")
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
+    parser.add_argument("--data-dir", type=Path, default=Path("../data"))
+    parser.add_argument("--gold-dir", type=Path, default=Path("goldensets/data"),
+                        help="Where goldenset_<cc>.jsonl live; used to enumerate annotated jurisdictions.")
+    parser.add_argument("--out-dir", type=Path, default=Path("inference-results/data"))
+    parser.add_argument("--prompt_version", default="v3")
+    args = parser.parse_args(argv)
+
+    written = 0
+    for gs in sorted(args.gold_dir.glob("*/goldenset_*.jsonl")):
+        cc = gs.parent.name
+        for src in sorted((args.data_dir / cc).glob(f"Goldenset_*_{args.prompt_version}_full_text_*.jsonl")):
+            m = _SLUG_RE.search(src.name)
+            model = SLUG_TO_MODEL.get(m.group(1)) if m else None
+            if model is None:
+                continue
+            dst = args.out_dir / cc / f"inference_{model}.jsonl"
+            dst.parent.mkdir(parents=True, exist_ok=True)
+            dst.write_text(src.read_text(encoding="utf-8"), encoding="utf-8")
+            written += 1
+    print(f"wrote {written} inference JSONL file(s) to {args.out_dir}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

build_legora_prompts_jsonl.py

@@ -0,0 +1,63 @@
+#!/usr/bin/env python3
+"""Build the published Legora prompt-metadata JSONL from the prompt-bearing export.
+
+The 2026-08-05 Legora re-export (``data/raw/legora_2026-08-05_prompts.xlsx``)
+carries the per-column question text of both runs in row 2: the left column
+group belongs to ``legora-1``, the right one to ``legora-2`` (see
+``scripts/compare_legora_prompts.py`` / ``data/analysis/legora_prompt_comparison.md``).
+This script writes one JSONL file per run with a record per field:
+
+    {"model": "legora-1", "inference_date": "2026-08-01", "field": ..., "prompt": ...}
+
+    python build_legora_prompts_jsonl.py --xlsx ../data/raw/legora_2026-08-05_prompts.xlsx \\
+        --out-dir inference-results/prompts
+"""
+import argparse
+import json
+from pathlib import Path
+
+import openpyxl
+
+RUNS = ("legora-1", "legora-2")
+INFERENCE_DATE = "2026-08-01"
+PROMPT_SUFFIX = " (with prompt)"
+
+
+def read_prompts(xlsx: Path) -> dict[str, dict[str, str]]:
+    """Return ``{model: {field: prompt}}`` from the export's prompt row."""
+    wb = openpyxl.load_workbook(xlsx, data_only=True, read_only=True)
+    ws = wb.worksheets[0]
+    rows = ws.iter_rows(min_row=1, max_row=2, values_only=True)
+    header, prompt_row = next(rows), next(rows)
+    out: dict[str, dict[str, str]] = {run: {} for run in RUNS}
+    for head, prompt in zip(header, prompt_row):
+        if not head or not str(head).endswith(PROMPT_SUFFIX) or not prompt:
+            continue
+        field = str(head)[: -len(PROMPT_SUFFIX)]
+        run = RUNS[0] if field not in out[RUNS[0]] else RUNS[1]
+        out[run][field] = " ".join(str(prompt).split())
+    return out
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
+    parser.add_argument("--xlsx", type=Path,
+                        default=Path("../data/raw/legora_2026-08-05_prompts.xlsx"))
+    parser.add_argument("--out-dir", type=Path, default=Path("inference-results/prompts"))
+    args = parser.parse_args(argv)
+
+    prompts = read_prompts(args.xlsx)
+    args.out_dir.mkdir(parents=True, exist_ok=True)
+    for run, fields in prompts.items():
+        dst = args.out_dir / f"prompts_{run.replace('-', '_')}.jsonl"
+        with dst.open("w", encoding="utf-8") as f:
+            for field, prompt in fields.items():
+                record = {"model": run, "inference_date": INFERENCE_DATE,
+                          "field": field, "prompt": prompt}
+                f.write(json.dumps(record, ensure_ascii=False) + "\n")
+        print(f"wrote {len(fields)} prompt record(s) to {dst}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

convert_goldenset_to_jsonl.py

@@ -2,11 +2,14 @@
 """Convert handcrafted Goldenset XLSX files to anonymised JSONL.
 
 For each ``Goldenset_*_final*.xlsx`` workbook under ``<data-dir>/<cc>/`` the
-GOLDENSET sheet is read, rows that the expert annotator did not fully
-classify are dropped (criterion: ``legal_subject_judgement`` must be
-populated, which the annotators used as the marker that a row has been
-substantively reviewed), and the remaining rows are written to
-``<out-dir>/<cc>/goldenset_<cc>.jsonl`` as one JSON object per line.
+GOLDENSET sheet is read, rows that the expert annotator did not review are
+dropped (criterion: at least one of the eleven evaluated label fields must
+carry a value — the same row-inclusion rule the scoring pipeline applies, so
+published-data scoring reproduces the paper), and the remaining rows are
+written to ``<out-dir>/<cc>/goldenset_<cc>.jsonl`` as one JSON object per
+line. A handful of reviewed rows carry no ``legal_subject_judgement``; they
+are excluded from the paper's judgment counts (which filter on that field)
+but included here because their cells enter the scoring denominators.
 
 Each output record contains the case identifiers (``case_id``, ``link``,
 ``full_text``) followed by the 14 schema fields defined in
@@ -24,10 +27,11 @@ Without arguments the script assumes ``./data`` for both inputs and outputs and
 processes the 19 jurisdictions of the paper.
 """
 
-from __future__ import annotations
-
 import argparse
+import hashlib
 import json
+import os
+import re
 import sys
 from datetime import date, datetime
 from pathlib import Path
@@ -57,8 +61,36 @@ SCHEMA_FIELDS = (
     "defendant_no1_ISIC1_industry_category",
 )
 
+# The evaluated label fields (currencies are scored via their amount field).
+# Row inclusion must match `legex.evaluation.scoring._read_goldenset_rows`.
+EVAL_FIELDS = tuple(f for f in SCHEMA_FIELDS if not f.startswith("Currency_"))
+
 EMPTY_LITERALS = frozenset({"", "none", "null", "nan"})
 
+# Loose case_id key so full_text.jsonl matches the workbook even when separators
+# differ (e.g. xlsx "4A_426/2024" vs jsonl "4A_426_2024").
+_CASE_ID_SEP_RE = re.compile(r"[\s/\\\-._;]+")
+
+
+def _norm_case_id(s: str) -> str:
+    return _CASE_ID_SEP_RE.sub("_", str(s).strip()).strip("_").lower()
+
+
+# Used to strip unnecessary prefixes
+_SOURCE_PREFIX_RE = re.compile(r"^case_id:.*?Quelle Text:\s*\S+\s+", re.DOTALL)
+
+
+def _strip_source_prefix(text):
+    if not text:
+        return text
+    return _SOURCE_PREFIX_RE.sub("", text, count=1)
+
+
+# Find failed scrapes.
+_SCRAPE_ERROR_RE = re.compile(
+    r"\(Fehlermeldung\)|Abruf fehlgeschlagen|Failed to resolve|Max retries exceeded"
+)
+
 
 def normalise(value: Any) -> Any:
     if value is None:
@@ -81,6 +113,74 @@ def normalise(value: Any) -> Any:
     return s
 
 
+# Field-type groups, so each JSONL column is single-typed
+_DATE_FIELDS = {"trial_start_date", "trial_end_date"}
+_INT_FIELDS = {"plaintiffs_all_count", "defendants_all_count"}
+_NUMBER_FIELDS = {
+    "plaintiff_loosing_share", "court_cost_awarded_nominal",
+    "party_compensation_awarded_nominal",
+} | _INT_FIELDS
+_ISIC_FIELDS = {
+    "plaintiff_no1_ISIC1_industry_category", "defendant_no1_ISIC1_industry_category",
+}
+_STRING_FIELDS = {
+    "legal_subject_judgement",
+    "Currency_dispute_value_nominal", "Currency_court_cost_awarded_nominal",
+    "Currency_party_compensation_awarded_nominal",
+}
+_DISPUTE_FIELD = "dispute_value_nominal"
+_ISO_DATE_RE = re.compile(r"^\d{4}-\d{2}-\d{2}$")
+
+
+def _to_number(value: Any) -> int | float | None:
+    if isinstance(value, bool):
+        return None
+    if isinstance(value, (int, float)):
+        f = float(value)
+        return int(f) if f.is_integer() else f
+    s = str(value).strip().replace("'", "").replace(" ", "").replace(" ", "")
+    if "_" in s:   # underscores here are range/typo separators, not digit grouping
+        return None
+    if re.fullmatch(r"-?\d{1,3}(\.\d{3})+,\d+", s):       # 1.000,50  (EU)
+        s = s.replace(".", "").replace(",", ".")
+    elif re.fullmatch(r"-?\d+,\d+", s):                   # 1234,56
+        s = s.replace(",", ".")
+    elif re.fullmatch(r"-?\d{1,3}(,\d{3})+(\.\d+)?", s):  # 1,000.50  (US)
+        s = s.replace(",", "")
+    try:
+        f = float(s)
+        if f != f or f in (float("inf"), float("-inf")):  # reject nan / inf
+            return None
+        return int(f) if f.is_integer() else f
+    except ValueError:
+        return None
+
+
+def normalise_field(field: str, value: Any) -> Any:
+    """Type-aware normalisation keeping each schema column single-typed."""
+    v = normalise(value)
+    if v is None:
+        return None
+    if field in _DATE_FIELDS:
+        s = str(v).strip()
+        return s if _ISO_DATE_RE.match(s) else None
+    if field == _DISPUTE_FIELD:
+        s = str(v).strip()
+        return "nonpecuniary" if s.lower() == "nonpecuniary" else s
+    if field in _ISIC_FIELDS:  # The categories are lower case
+        return str(v).strip().lower()
+    if field in _NUMBER_FIELDS:
+        n = _to_number(v)
+        if n is None:
+            return None
+        # counts -> int; money/ratio -> natural form (int when whole, else float).
+        # pyarrow promotes int/float within a split, so the viewer renders fine.
+        return int(n) if field in _INT_FIELDS else n
+    if field in _STRING_FIELDS:
+        return str(v).strip()
+    return v
+
+
 def find_goldenset_xlsx(data_dir: Path, cc: str) -> Path | None:
     """Return the *_final*.xlsx workbook for a jurisdiction, if any."""
     jurisdiction_dir = data_dir / cc
@@ -104,6 +204,7 @@ def load_full_text_fallback(data_dir: Path, cc: str) -> dict[str, str]:
     if not path.exists():
         return {}
     fallback: dict[str, str] = {}
+    norm_extra: dict[str, str] = {}
     for line in path.read_text(encoding="utf-8").splitlines():
         if not line.strip():
             continue
@@ -112,6 +213,9 @@ def load_full_text_fallback(data_dir: Path, cc: str) -> dict[str, str]:
         text = record.get("full_text") or record.get("text")
         if case_id and text:
             fallback[str(case_id)] = str(text)
+            norm_extra.setdefault(_norm_case_id(case_id), str(text))
+    for k, v in norm_extra.items():  # normalized keys never overwrite exact ones
+        fallback.setdefault(k, v)
     return fallback
 
 
@@ -132,18 +236,30 @@ def convert_workbook(xlsx_path: Path, fallback: dict[str, str]) -> list[dict[str
         case_id = normalise(cells.get("case_id"))
         if not case_id:
             continue
-        labels = {field: normalise(cells.get(field)) for field in SCHEMA_FIELDS}
-        if labels["legal_subject_judgement"] is None:
+        # Same row-inclusion rule as the scoring pipeline: reviewed = at least one
+        # evaluated field carries a value in the workbook (raw, pre-typing — a non-ISO
+        # date string still marks the row as reviewed even though it publishes as null).
+        if all(normalise(cells.get(field)) is None for field in EVAL_FIELDS):
             continue
+        labels = {field: normalise_field(field, cells.get(field)) for field in SCHEMA_FIELDS}
         full_text = normalise(cells.get("full_text"))
+        if full_text and _SCRAPE_ERROR_RE.search(full_text):
+            full_text = None  # Broken scrape
         if not full_text:
-            full_text = fallback.get(str(case_id))
+            full_text = fallback.get(str(case_id)) or fallback.get(_norm_case_id(case_id))
+        full_text = _strip_source_prefix(full_text)
         record: dict[str, Any] = {
             "case_id": str(case_id),
             "link": normalise(cells.get("link")),
             "full_text": full_text,
         }
         record.update(labels)
+        # Provenance columns (workbook-only otherwise): included on request so the
+        # published JSONL carries the sanitization audit trail. original_input is a
+        # JSON object of the pre-sanitization values for corrected fields; rows that
+        # needed no correction get an empty object "{}" (never null).
+        record["comment"] = normalise(cells.get("comment"))
+        record["original_input"] = normalise(cells.get("original_input")) or "{}"
         records.append(record)
     return records
 
@@ -155,6 +271,24 @@ def write_jsonl(records: list[dict[str, Any]], out_path: Path) -> None:
             f.write(json.dumps(record, ensure_ascii=False) + "\n")
 
 
+def _load_annotators(data_dir: Path) -> dict | None:
+    """Local annotator lookup (gitignored): {primary: {cc: name}, reannotations: [...]}.
+
+    Names never reach the published JSONL — only the salted-hash annotator_id does.
+    """
+    path = data_dir / "reannotation" / "annotators.json"
+    if not path.exists():
+        return None
+    return json.loads(path.read_text(encoding="utf-8"))
+
+
+def _annotator_id(name: str | None, salt: str) -> str | None:
+    """Pseudonymous, stable per-name id (same name -> same id across countries)."""
+    if not name:
+        return None
+    return hashlib.sha256(f"{salt}|{name}".encode("utf-8")).hexdigest()[:10]
+
+
 def main(argv: list[str] | None = None) -> int:
     parser = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
     parser.add_argument(
@@ -185,6 +319,16 @@ def main(argv: list[str] | None = None) -> int:
     data_dir: Path = args.data_dir.resolve()
     out_dir: Path = args.out_dir.resolve()
 
+    try:  # pick up ANNOTATOR_SALT from .env when available
+        from dotenv import load_dotenv
+        load_dotenv()
+    except ImportError:
+        pass
+    annotators = _load_annotators(data_dir)
+    salt = os.environ.get("ANNOTATOR_SALT", "")
+    if annotators and not salt:
+        print("warning: annotators.json found but ANNOTATOR_SALT is unset", file=sys.stderr)
+
     total = 0
     missing: list[str] = []
     for cc in args.jurisdictions:
@@ -195,10 +339,29 @@ def main(argv: list[str] | None = None) -> int:
             continue
         fallback = load_full_text_fallback(data_dir, cc)
         records = convert_workbook(xlsx, fallback)
+        n_reann = 0
+        if annotators:  # stamp annotator_id; append second-annotator rows
+            pid = _annotator_id(annotators.get("primary", {}).get(cc), salt)
+            for r in records:
+                r["annotator_id"] = pid
+            for entry in annotators.get("reannotations", []):
+                if entry.get("cc") != cc:
+                    continue
+                rxlsx = data_dir / entry["file"]
+                if not rxlsx.exists():
+                    print(f"[{cc}] reannotation missing: {rxlsx}", file=sys.stderr)
+                    continue
+                rid = _annotator_id(entry["name"], salt)
+                rrecords = convert_workbook(rxlsx, fallback)
+                for r in rrecords:
+                    r["annotator_id"] = rid
+                records.extend(rrecords)
+                n_reann += len(rrecords)
         out_path = out_dir / cc / f"goldenset_{cc}.jsonl"
         if not args.dry_run:
             write_jsonl(records, out_path)
-        print(f"[{cc}] {xlsx.name} -> {out_path.relative_to(out_dir.parent)}: {len(records)} rows")
+        extra = f" (+{n_reann} reann)" if n_reann else ""
+        print(f"[{cc}] {xlsx.name} -> {out_path.relative_to(out_dir.parent)}: {len(records)} rows{extra}")
         total += len(records)
 
     print(f"\nTotal: {total} rows across {len(args.jurisdictions) - len(missing)} jurisdictions.")

data/analysis/hallucinations/hallucination_review_ch.csv

@@ -0,0 +1,197 @@
+case_id,field,model,category
+4A_211/2024,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+4A_211/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4A_211/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,B
+4A_211/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,A
+4A_211/2024,trial_start_date,gpt-5.4-mini,B
+4A_312/2024,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+4A_312/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4A_312/2024,defendant_no1_ISIC1_industry_category,harvey,D
+4A_312/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+4A_312/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4A_312/2024,plaintiff_no1_ISIC1_industry_category,harvey,D
+4A_312/2024,trial_start_date,gemini/gemini-3.1-flash-lite,B
+4A_312/2024,trial_start_date,gpt-5.4-mini,B
+4A_317/2024,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,A
+4A_317/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,A
+4A_317/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+4A_317/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4A_317/2024,trial_start_date,gpt-5.4-mini,B
+4A_317/2024,trial_start_date,harvey,E
+4A_372/2023,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,B
+4A_372/2023,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4A_372/2023,defendant_no1_ISIC1_industry_category,harvey,D
+4A_372/2023,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+4A_372/2023,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4A_372/2023,plaintiff_no1_ISIC1_industry_category,harvey,C
+4A_377/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4A_377/2024,defendant_no1_ISIC1_industry_category,harvey,D
+4A_377/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4A_377/2024,dispute_value_nominal,gemini/gemini-3.1-flash-lite,B
+4A_377/2024,dispute_value_nominal,gpt-5.4-mini,B
+4A_377/2024,trial_start_date,gpt-5.4-mini,B
+4A_559/2025,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4A_559/2025,defendant_no1_ISIC1_industry_category,harvey,A
+4A_559/2025,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4A_559/2025,dispute_value_nominal,gemini/gemini-3.1-flash-lite,A
+4A_559/2025,dispute_value_nominal,gpt-5.4-mini,A
+4A_606/2024,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+4A_606/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4A_606/2024,defendant_no1_ISIC1_industry_category,harvey,D
+4A_606/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,B
+4A_606/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,C
+4A_606/2024,plaintiff_no1_ISIC1_industry_category,harvey,C
+4A_606/2024,trial_start_date,gpt-5.4-mini,B
+4A_608/2023,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,A
+4A_608/2023,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,A
+4A_608/2023,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,A
+4A_608/2023,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,A
+4A_608/2023,dispute_value_nominal,gpt-5.4-mini,A
+4A_608/2023,trial_start_date,gpt-5.4-mini,B
+4D_115/2024,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+4D_115/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4D_115/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+4D_115/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4D_115/2024,plaintiff_no1_ISIC1_industry_category,harvey,D
+4D_147/2024,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,B
+4D_147/2024,defendant_no1_ISIC1_industry_category,harvey,B
+4D_147/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4D_147/2024,trial_start_date,gemini/gemini-3.1-flash-lite,B
+4D_147/2024,trial_start_date,gpt-5.4-mini,B
+4D_179/2025,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,B
+4D_179/2025,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4D_179/2025,defendant_no1_ISIC1_industry_category,harvey,C
+4D_179/2025,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+4D_179/2025,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4D_179/2025,trial_start_date,gpt-5.4-mini,B
+4D_179/2025,trial_start_date,harvey,E
+4D_40/2025,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+4D_40/2025,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+4D_40/2025,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,A
+4D_40/2025,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,A
+4D_40/2025,dispute_value_nominal,gemini/gemini-3.1-flash-lite,C
+4D_40/2025,dispute_value_nominal,gpt-5.4-mini,C
+4D_40/2025,dispute_value_nominal,harvey,C
+5A_107/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,A
+5A_107/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,A
+5A_107/2024,party_compensation_awarded_nominal,gpt-5.4-mini,A
+5A_107/2024,dispute_value_nominal,gemini/gemini-3.1-flash-lite,C
+5A_107/2024,dispute_value_nominal,gpt-5.4-mini,C
+5A_107/2024,dispute_value_nominal,harvey,C
+5A_119/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_119/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_119/2024,dispute_value_nominal,gemini/gemini-3.1-flash-lite,D
+5A_119/2024,dispute_value_nominal,gpt-5.4-mini,D
+5A_119/2024,dispute_value_nominal,harvey,D
+5A_140/2025,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_140/2025,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_140/2025,party_compensation_awarded_nominal,gpt-5.4-mini,A
+5A_140/2025,dispute_value_nominal,gemini/gemini-3.1-flash-lite,A
+5A_140/2025,dispute_value_nominal,gpt-5.4-mini,A
+5A_169/2024,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+5A_169/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_169/2024,defendant_no1_ISIC1_industry_category,harvey,D
+5A_169/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+5A_169/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_169/2024,party_compensation_awarded_nominal,gemini/gemini-3.1-flash-lite,A
+5A_169/2024,party_compensation_awarded_nominal,gpt-5.4-mini,C
+5A_169/2024,party_compensation_awarded_nominal,harvey,C
+5A_190/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+5A_190/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_190/2024,plaintiff_no1_ISIC1_industry_category,harvey,D
+5A_190/2024,party_compensation_awarded_nominal,gemini/gemini-3.1-flash-lite,A
+5A_190/2024,party_compensation_awarded_nominal,gpt-5.4-mini,A
+5A_220/2024,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,C
+5A_220/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,C
+5A_220/2024,defendant_no1_ISIC1_industry_category,harvey,C
+5A_220/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+5A_220/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_220/2024,party_compensation_awarded_nominal,gpt-5.4-mini,A
+5A_220/2024,dispute_value_nominal,gemini/gemini-3.1-flash-lite,A
+5A_341/2023,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_341/2023,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,A
+5A_341/2023,dispute_value_nominal,gemini/gemini-3.1-flash-lite,B
+5A_341/2023,dispute_value_nominal,gpt-5.4-mini,B
+5A_341/2023,dispute_value_nominal,harvey,B
+5A_358/2024,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,A
+5A_358/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,A
+5A_358/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,C
+5A_358/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,C
+5A_358/2024,party_compensation_awarded_nominal,gpt-5.4-mini,A
+5A_358/2024,dispute_value_nominal,gemini/gemini-3.1-flash-lite,C
+5A_358/2024,dispute_value_nominal,gpt-5.4-mini,C
+5A_358/2024,dispute_value_nominal,harvey,C
+5A_38/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+5A_38/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_38/2024,party_compensation_awarded_nominal,gemini/gemini-3.1-flash-lite,D
+5A_38/2024,party_compensation_awarded_nominal,gpt-5.4-mini,D
+5A_38/2024,dispute_value_nominal,gpt-5.4-mini,A
+5A_389/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+5A_389/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_389/2024,party_compensation_awarded_nominal,gpt-5.4-mini,A
+5A_389/2024,dispute_value_nominal,gemini/gemini-3.1-flash-lite,A
+5A_389/2024,dispute_value_nominal,gpt-5.4-mini,A
+5A_47/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+5A_47/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_47/2024,plaintiff_no1_ISIC1_industry_category,harvey,A
+5A_47/2024,party_compensation_awarded_nominal,gpt-5.4-mini,B
+5A_47/2024,dispute_value_nominal,gemini/gemini-3.1-flash-lite,C
+5A_47/2024,dispute_value_nominal,gpt-5.4-mini,C
+5A_47/2024,dispute_value_nominal,harvey,C
+5A_499/2024,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+5A_499/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_499/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+5A_499/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_499/2024,party_compensation_awarded_nominal,gpt-5.4-mini,A
+5A_503/2023,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_503/2023,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_503/2023,plaintiff_no1_ISIC1_industry_category,harvey,D
+5A_503/2023,dispute_value_nominal,gemini/gemini-3.1-flash-lite,B
+5A_503/2023,dispute_value_nominal,gpt-5.4-mini,A
+5A_592/2025,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_592/2025,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_592/2025,party_compensation_awarded_nominal,gpt-5.4-mini,A
+5A_592/2025,dispute_value_nominal,gemini/gemini-3.1-flash-lite,A
+5A_592/2025,dispute_value_nominal,gpt-5.4-mini,A
+5A_597/2023,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+5A_597/2023,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_597/2023,plaintiff_no1_ISIC1_industry_category,harvey,D
+5A_597/2023,dispute_value_nominal,gemini/gemini-3.1-flash-lite,B
+5A_597/2023,dispute_value_nominal,gpt-5.4-mini,B
+5A_679/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,D
+5A_679/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5A_679/2024,dispute_value_nominal,gemini/gemini-3.1-flash-lite,A
+5A_679/2024,dispute_value_nominal,harvey,C
+5A_679/2024,plaintiff_loosing_share,gpt-5.4-mini,A
+5A_891/2024,defendant_no1_ISIC1_industry_category,harvey,B
+5A_891/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,C
+5A_891/2024,party_compensation_awarded_nominal,gpt-5.4-mini,D
+5A_891/2024,plaintiff_loosing_share,gemini/gemini-3.1-flash-lite,C
+5A_891/2024,plaintiff_loosing_share,gpt-5.4-mini,C
+5A_891/2024,plaintiff_loosing_share,harvey,C
+5D_63/2024,defendant_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,B
+5D_63/2024,defendant_no1_ISIC1_industry_category,gpt-5.4-mini,A
+5D_63/2024,defendant_no1_ISIC1_industry_category,harvey,C
+5D_63/2024,plaintiff_no1_ISIC1_industry_category,gemini/gemini-3.1-flash-lite,A
+5D_63/2024,plaintiff_no1_ISIC1_industry_category,gpt-5.4-mini,D
+5D_63/2024,plaintiff_no1_ISIC1_industry_category,harvey,D
+5D_63/2024,dispute_value_nominal,gemini/gemini-3.1-flash-lite,C
+5D_63/2024,dispute_value_nominal,gpt-5.4-mini,C
+5D_63/2024,dispute_value_nominal,harvey,C
+4A_211/2024,defendant_no1_ISIC1_industry_category,legora-1,D
+4A_372/2023,plaintiff_no1_ISIC1_industry_category,legora-1,C
+4A_559/2025,dispute_value_nominal,legora-1,A
+4A_606/2024,plaintiff_no1_ISIC1_industry_category,legora-1,C
+4A_608/2023,dispute_value_nominal,legora-1,A
+4D_115/2024,defendant_no1_ISIC1_industry_category,legora-1,D
+4D_40/2025,dispute_value_nominal,legora-1,C
+5A_107/2024,dispute_value_nominal,legora-1,C
+5A_119/2024,dispute_value_nominal,legora-1,D
+5A_169/2024,party_compensation_awarded_nominal,legora-1,C
+5A_341/2023,dispute_value_nominal,legora-1,B
+5A_358/2024,dispute_value_nominal,legora-1,C
+5A_47/2024,dispute_value_nominal,legora-1,C
+5A_503/2023,plaintiff_no1_ISIC1_industry_category,legora-1,D
+5A_891/2024,plaintiff_loosing_share,legora-1,C
+5D_63/2024,defendant_no1_ISIC1_industry_category,legora-1,B
+5D_63/2024,dispute_value_nominal,legora-1,C

data/analysis/iaa/ANALYSIS.md

@@ -0,0 +1,362 @@
+# 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`.

data/analysis/iaa/alt_test_decomposition.csv

@@ -0,0 +1,77 @@
+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
+gpt-5.4-mini,br,all,570,185,72,53,445,401,44,283,123,39,0.907,0.5167
+gpt-5.4-mini,br,nontrivial,320,185,52,22,246,208,38,97,123,26,0.9313,0.5469
+gpt-5.4-mini,ch,all,900,153,46,95,759,704,55,628,90,41,0.8944,0.4728
+gpt-5.4-mini,ch,nontrivial,736,153,46,26,664,615,49,550,90,24,0.9647,0.5136
+gpt-5.4-mini,de,all,600,127,34,114,452,394,58,326,74,52,0.81,0.4333
+gpt-5.4-mini,de,nontrivial,397,127,25,87,285,230,55,173,74,38,0.7809,0.4219
+gpt-5.4-mini,ge,all,570,238,75,104,391,281,110,153,153,85,0.8175,0.4746
+gpt-5.4-mini,ge,nontrivial,490,238,68,85,337,251,86,120,153,64,0.8265,0.4827
+gpt-5.4-mini,sg,all,600,233,50,83,467,350,117,250,119,98,0.8617,0.4725
+gpt-5.4-mini,sg,nontrivial,498,233,39,74,385,288,97,183,119,83,0.8514,0.4649
+gpt-5.4-mini,tw,all,600,129,34,83,483,419,64,347,92,44,0.8617,0.4592
+gpt-5.4-mini,tw,nontrivial,416,129,32,53,331,270,61,216,92,23,0.8726,0.4748
+gpt-5.4-mini,uk,all,900,337,83,169,648,502,146,358,156,134,0.8122,0.4522
+gpt-5.4-mini,uk,nontrivial,671,337,74,142,455,319,136,177,156,122,0.7884,0.4493
+gpt-5.4-mini,us,all,900,141,60,198,642,603,39,543,48,51,0.78,0.4233
+gpt-5.4-mini,us,nontrivial,688,141,52,186,450,412,38,360,48,42,0.7297,0.4026
+gemini/gemini-3.1-flash-lite,br,all,570,185,74,33,463,422,41,303,123,37,0.9421,0.536
+gemini/gemini-3.1-flash-lite,br,nontrivial,320,185,54,26,240,206,34,93,123,24,0.9187,0.5437
+gemini/gemini-3.1-flash-lite,ch,all,900,153,53,89,758,721,37,638,86,34,0.9011,0.48
+gemini/gemini-3.1-flash-lite,ch,nontrivial,736,153,46,53,637,605,32,527,86,24,0.928,0.4952
+gemini/gemini-3.1-flash-lite,de,all,600,127,37,93,470,416,54,347,74,49,0.845,0.4533
+gemini/gemini-3.1-flash-lite,de,nontrivial,397,127,30,81,286,235,51,179,74,33,0.796,0.4358
+gemini/gemini-3.1-flash-lite,ge,all,570,238,65,78,427,310,117,165,177,85,0.8632,0.4886
+gemini/gemini-3.1-flash-lite,ge,nontrivial,490,238,53,69,368,273,95,122,177,69,0.8592,0.4837
+gemini/gemini-3.1-flash-lite,sg,all,600,233,68,67,465,371,94,250,135,80,0.8883,0.5008
+gemini/gemini-3.1-flash-lite,sg,nontrivial,498,233,50,64,384,304,80,177,135,72,0.8715,0.4859
+gemini/gemini-3.1-flash-lite,tw,all,600,129,38,74,488,432,56,356,92,40,0.8767,0.47
+gemini/gemini-3.1-flash-lite,tw,nontrivial,416,129,29,56,331,278,53,213,92,26,0.8654,0.4675
+gemini/gemini-3.1-flash-lite,uk,all,900,337,87,163,650,508,142,352,168,130,0.8189,0.4578
+gemini/gemini-3.1-flash-lite,uk,nontrivial,671,337,74,145,452,316,136,162,168,122,0.7839,0.4471
+gemini/gemini-3.1-flash-lite,us,all,900,141,59,224,617,577,40,519,46,52,0.7511,0.4083
+gemini/gemini-3.1-flash-lite,us,nontrivial,688,141,52,218,418,381,37,330,46,42,0.6831,0.3794
+harvey,br,all,570,185,72,47,451,405,46,289,121,41,0.9175,0.5219
+harvey,br,nontrivial,320,185,51,43,226,187,39,76,121,29,0.8656,0.5125
+harvey,ch,all,900,153,59,61,780,747,33,656,96,28,0.9322,0.4989
+harvey,ch,nontrivial,736,153,49,55,632,601,31,515,96,21,0.9253,0.4959
+harvey,de,all,600,127,46,99,455,421,34,341,74,40,0.835,0.4558
+harvey,de,nontrivial,397,127,28,93,276,246,30,167,74,35,0.7657,0.4181
+harvey,ge,all,570,238,79,92,399,296,103,151,175,73,0.8386,0.4886
+harvey,ge,nontrivial,490,238,56,89,345,263,82,102,175,68,0.8184,0.4663
+harvey,sg,all,600,233,78,49,473,391,82,264,139,70,0.9183,0.5242
+harvey,sg,nontrivial,498,233,57,49,392,324,68,188,139,65,0.9016,0.508
+harvey,tw,all,600,129,55,79,466,440,26,347,96,23,0.8683,0.48
+harvey,tw,nontrivial,416,129,35,70,311,288,23,195,96,20,0.8317,0.4579
+harvey,uk,all,900,337,139,68,693,615,78,398,217,78,0.9244,0.5394
+harvey,uk,nontrivial,671,337,124,56,491,417,74,202,217,72,0.9165,0.5507
+harvey,us,all,900,141,66,195,639,611,28,540,54,45,0.7833,0.4283
+harvey,us,nontrivial,688,141,51,186,451,424,27,354,54,43,0.7297,0.4019
+legora-1,br,all,540,174,68,38,434,395,39,275,125,34,0.9296,0.5278
+legora-1,br,nontrivial,306,174,40,35,231,196,35,73,125,33,0.8856,0.5082
+legora-1,ch,all,900,153,59,23,818,783,35,684,106,28,0.9744,0.52
+legora-1,ch,nontrivial,736,153,45,23,668,633,35,537,106,25,0.9688,0.5149
+legora-1,de,all,570,123,49,169,352,323,29,242,76,34,0.7035,0.3947
+legora-1,de,nontrivial,378,123,29,169,180,151,29,71,76,33,0.5529,0.3148
+legora-1,sg,all,600,233,74,94,432,350,82,229,129,74,0.8433,0.4833
+legora-1,sg,nontrivial,498,233,49,94,355,287,68,153,129,73,0.8112,0.4548
+legora-1,tw,all,600,129,57,63,480,457,23,359,100,21,0.895,0.495
+legora-1,tw,nontrivial,416,129,35,57,324,303,21,204,100,20,0.863,0.4736
+legora-1,uk,all,900,337,128,93,679,590,89,382,208,89,0.8967,0.5194
+legora-1,uk,nontrivial,671,337,109,93,469,382,87,174,208,87,0.8614,0.5119
+legora-1,us,all,900,141,71,269,560,539,21,468,52,40,0.7011,0.39
+legora-1,us,nontrivial,688,141,58,266,364,344,20,276,52,36,0.6134,0.3488
+legora-2,br,all,510,166,65,37,408,370,38,262,111,35,0.9275,0.5275
+legora-2,br,nontrivial,297,166,41,33,223,188,35,79,111,33,0.8889,0.5135
+legora-2,ch,all,900,153,62,34,804,774,30,679,100,25,0.9622,0.5156
+legora-2,ch,nontrivial,736,153,46,34,656,627,29,532,100,24,0.9538,0.5082
+legora-2,de,all,600,127,50,188,362,332,30,248,78,36,0.6867,0.385
+legora-2,de,nontrivial,397,127,29,188,180,150,30,68,78,34,0.5264,0.2997
+legora-2,sg,all,600,233,72,88,440,357,83,241,123,76,0.8533,0.4867
+legora-2,sg,nontrivial,498,233,47,88,363,294,69,165,123,75,0.8233,0.4588
+legora-2,tw,all,600,129,56,60,484,460,24,362,100,22,0.9,0.4967
+legora-2,tw,nontrivial,416,129,33,54,329,307,22,207,100,22,0.8702,0.4748
+legora-2,uk,all,900,337,131,83,686,600,86,390,210,86,0.9078,0.5267
+legora-2,uk,nontrivial,671,337,111,82,478,393,85,183,210,85,0.8778,0.5216
+legora-2,us,all,900,141,72,258,570,553,17,483,48,39,0.7133,0.3967
+legora-2,us,nontrivial,688,141,57,255,376,361,15,291,48,37,0.6294,0.3561

data/analysis/iaa/alt_test_pooled.csv

@@ -0,0 +1,39 @@
+candidate,country,omega,rho,passes,omega_nontrivial,rho_nontrivial,passes_nontrivial
+gpt-5.4-mini,br,1.0,0.9,1,1.0,0.9352,1
+gpt-5.4-mini,ch,1.0,0.8944,1,1.0,0.9655,1
+gpt-5.4-mini,de,0.3333,0.7961,0,0.3333,0.7783,0
+gpt-5.4-mini,ge,1.0,0.8189,1,1.0,0.8417,1
+gpt-5.4-mini,sg,1.0,0.8583,1,1.0,0.8698,1
+gpt-5.4-mini,tw,1.0,0.8561,1,1.0,0.8703,1
+gpt-5.4-mini,uk,1.0,0.8122,1,1.0,0.7951,1
+gpt-5.4-mini,us,0.0,0.78,0,0.0,0.7362,0
+gemini/gemini-3.1-flash-lite,br,1.0,0.934,1,1.0,0.9207,1
+gemini/gemini-3.1-flash-lite,ch,1.0,0.9011,1,1.0,0.9296,1
+gemini/gemini-3.1-flash-lite,de,1.0,0.8433,1,0.3333,0.7992,0
+gemini/gemini-3.1-flash-lite,ge,1.0,0.8611,1,1.0,0.8695,1
+gemini/gemini-3.1-flash-lite,sg,1.0,0.8815,1,1.0,0.8977,1
+gemini/gemini-3.1-flash-lite,tw,1.0,0.8633,1,1.0,0.8565,1
+gemini/gemini-3.1-flash-lite,uk,1.0,0.8189,1,1.0,0.7908,1
+gemini/gemini-3.1-flash-lite,us,0.0,0.7511,0,0.0,0.6908,0
+harvey,br,1.0,0.9108,1,1.0,0.8697,1
+harvey,ch,1.0,0.9322,1,1.0,0.927,1
+harvey,de,0.6667,0.8272,1,0.6667,0.7678,1
+harvey,ge,1.0,0.8355,1,1.0,0.8358,1
+harvey,sg,1.0,0.9227,1,1.0,0.9149,1
+harvey,tw,1.0,0.8461,1,1.0,0.8157,1
+harvey,uk,1.0,0.9244,1,1.0,0.9192,1
+harvey,us,0.3333,0.7833,0,0.0,0.7362,0
+legora-1,br,1.0,0.9143,1,1.0,0.8685,1
+legora-1,ch,1.0,0.9744,1,1.0,0.9695,1
+legora-1,de,0.0,0.6993,0,0.0,0.5682,0
+legora-1,sg,1.0,0.8454,1,1.0,0.8241,1
+legora-1,tw,1.0,0.8783,1,1.0,0.8501,1
+legora-1,uk,1.0,0.8967,1,1.0,0.8658,1
+legora-1,us,0.0,0.7011,0,0.0,0.6227,0
+legora-2,br,1.0,0.915,1,1.0,0.8719,1
+legora-2,ch,1.0,0.9622,1,1.0,0.9548,1
+legora-2,de,0.0,0.6861,0,0.0,0.5492,0
+legora-2,sg,1.0,0.8582,1,1.0,0.8387,1
+legora-2,tw,1.0,0.8778,1,1.0,0.8524,1
+legora-2,uk,1.0,0.9078,1,1.0,0.8802,1
+legora-2,us,0.0,0.7133,0,0.0,0.6383,0

data/analysis/iaa/alt_test_reference_gemini_gemini-3.1-flash-lite.csv

@@ -0,0 +1,81 @@
+candidate,country,field,winning_rate,advantage_probability,passes,winning_rate_nontrivial,advantage_probability_nontrivial,passes_nontrivial
+gemini/gemini-3.1-flash-lite,br,court_cost_awarded_nominal,1.0,1.0,1,,,
+gemini/gemini-3.1-flash-lite,br,defendant_no1_ISIC1_industry_category,0.6667,0.8474,1,0.6667,0.8829,1
+gemini/gemini-3.1-flash-lite,br,defendants_all_count,1.0,0.9556,1,1.0,0.9556,1
+gemini/gemini-3.1-flash-lite,br,dispute_value_nominal,1.0,1.0,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,br,party_compensation_awarded_nominal,1.0,0.9778,1,,,
+gemini/gemini-3.1-flash-lite,br,plaintiff_loosing_share,0.0,0.8538,0,0.0,0.8194,0
+gemini/gemini-3.1-flash-lite,br,plaintiff_no1_ISIC1_industry_category,0.6667,0.8345,1,0.6667,0.8317,1
+gemini/gemini-3.1-flash-lite,br,plaintiffs_all_count,1.0,0.9778,1,1.0,0.9778,1
+gemini/gemini-3.1-flash-lite,br,trial_end_date,1.0,0.938,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,br,trial_start_date,1.0,0.9556,1,,,
+gemini/gemini-3.1-flash-lite,ch,court_cost_awarded_nominal,1.0,0.9778,1,1.0,0.9778,1
+gemini/gemini-3.1-flash-lite,ch,defendant_no1_ISIC1_industry_category,0.3333,0.8889,0,0.0,0.8611,0
+gemini/gemini-3.1-flash-lite,ch,defendants_all_count,1.0,0.9778,1,1.0,0.9778,1
+gemini/gemini-3.1-flash-lite,ch,dispute_value_nominal,1.0,0.9,1,1.0,0.9444,1
+gemini/gemini-3.1-flash-lite,ch,party_compensation_awarded_nominal,1.0,0.9778,1,1.0,0.9649,1
+gemini/gemini-3.1-flash-lite,ch,plaintiff_loosing_share,1.0,0.9556,1,1.0,0.9556,1
+gemini/gemini-3.1-flash-lite,ch,plaintiff_no1_ISIC1_industry_category,0.0,0.6111,0,0.0,0.7708,0
+gemini/gemini-3.1-flash-lite,ch,plaintiffs_all_count,1.0,1.0,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,ch,trial_end_date,1.0,1.0,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,ch,trial_start_date,0.0,0.7222,0,0.0,0.7361,0
+gemini/gemini-3.1-flash-lite,de,court_cost_awarded_nominal,1.0,1.0,1,,,
+gemini/gemini-3.1-flash-lite,de,defendant_no1_ISIC1_industry_category,0.3333,0.8,0,0.3333,0.8528,0
+gemini/gemini-3.1-flash-lite,de,defendants_all_count,1.0,0.9778,1,1.0,0.9778,1
+gemini/gemini-3.1-flash-lite,de,dispute_value_nominal,1.0,1.0,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,de,party_compensation_awarded_nominal,1.0,1.0,1,,,
+gemini/gemini-3.1-flash-lite,de,plaintiff_loosing_share,0.0,0.7278,0,0.0,0.7568,0
+gemini/gemini-3.1-flash-lite,de,plaintiff_no1_ISIC1_industry_category,0.6667,0.9444,1,0.6667,0.9437,1
+gemini/gemini-3.1-flash-lite,de,plaintiffs_all_count,0.6667,0.9278,1,0.6667,0.9278,1
+gemini/gemini-3.1-flash-lite,de,trial_end_date,0.0,0.1333,0,0.0,0.1333,0
+gemini/gemini-3.1-flash-lite,de,trial_start_date,0.6667,0.9222,1,,,
+gemini/gemini-3.1-flash-lite,ge,court_cost_awarded_nominal,0.3333,0.7778,0,0.3333,0.7644,0
+gemini/gemini-3.1-flash-lite,ge,defendant_no1_ISIC1_industry_category,0.3333,0.8429,0,0.6667,0.8889,1
+gemini/gemini-3.1-flash-lite,ge,defendants_all_count,1.0,1.0,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,ge,dispute_value_nominal,0.0,0.6973,0,0.0,0.7444,0
+gemini/gemini-3.1-flash-lite,ge,party_compensation_awarded_nominal,1.0,1.0,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,ge,plaintiff_loosing_share,1.0,0.8874,1,1.0,0.8874,1
+gemini/gemini-3.1-flash-lite,ge,plaintiff_no1_ISIC1_industry_category,0.0,0.7918,0,0.0,0.8245,0
+gemini/gemini-3.1-flash-lite,ge,plaintiffs_all_count,1.0,1.0,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,ge,trial_end_date,0.0,0.6822,0,0.0,0.6822,0
+gemini/gemini-3.1-flash-lite,ge,trial_start_date,1.0,0.9322,1,0.6667,0.9183,1
+gemini/gemini-3.1-flash-lite,sg,court_cost_awarded_nominal,1.0,0.948,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,sg,defendant_no1_ISIC1_industry_category,0.3333,0.8276,0,1.0,0.8731,1
+gemini/gemini-3.1-flash-lite,sg,defendants_all_count,1.0,0.9599,1,1.0,0.9599,1
+gemini/gemini-3.1-flash-lite,sg,dispute_value_nominal,0.0,0.8157,0,0.0,0.8449,0
+gemini/gemini-3.1-flash-lite,sg,party_compensation_awarded_nominal,1.0,0.948,1,1.0,0.963,1
+gemini/gemini-3.1-flash-lite,sg,plaintiff_loosing_share,0.0,0.7522,0,0.0,0.7722,0
+gemini/gemini-3.1-flash-lite,sg,plaintiff_no1_ISIC1_industry_category,0.3333,0.7844,0,0.3333,0.8401,0
+gemini/gemini-3.1-flash-lite,sg,plaintiffs_all_count,1.0,1.0,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,sg,trial_end_date,1.0,1.0,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,sg,trial_start_date,0.3333,0.7791,0,0.3333,0.7728,0
+gemini/gemini-3.1-flash-lite,tw,court_cost_awarded_nominal,0.6667,0.8889,1,0.0,0.7059,0
+gemini/gemini-3.1-flash-lite,tw,defendant_no1_ISIC1_industry_category,0.3333,0.8556,0,0.6667,0.8899,1
+gemini/gemini-3.1-flash-lite,tw,defendants_all_count,1.0,0.9778,1,1.0,0.9778,1
+gemini/gemini-3.1-flash-lite,tw,dispute_value_nominal,0.0,0.8,0,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,tw,party_compensation_awarded_nominal,1.0,1.0,1,,,
+gemini/gemini-3.1-flash-lite,tw,plaintiff_loosing_share,0.0,0.8444,0,0.0,0.8141,0
+gemini/gemini-3.1-flash-lite,tw,plaintiff_no1_ISIC1_industry_category,0.6667,0.9333,1,1.0,0.971,1
+gemini/gemini-3.1-flash-lite,tw,plaintiffs_all_count,1.0,0.9778,1,1.0,0.9778,1
+gemini/gemini-3.1-flash-lite,tw,trial_end_date,1.0,1.0,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,tw,trial_start_date,0.0,0.3556,0,0.0,0.2,0
+gemini/gemini-3.1-flash-lite,uk,court_cost_awarded_nominal,1.0,1.0,1,,,
+gemini/gemini-3.1-flash-lite,uk,defendant_no1_ISIC1_industry_category,0.0,0.7222,0,0.0,0.7222,0
+gemini/gemini-3.1-flash-lite,uk,defendants_all_count,0.0,0.8333,0,0.0,0.8333,0
+gemini/gemini-3.1-flash-lite,uk,dispute_value_nominal,0.3333,0.8444,0,1.0,0.8889,1
+gemini/gemini-3.1-flash-lite,uk,party_compensation_awarded_nominal,1.0,0.9333,1,,,
+gemini/gemini-3.1-flash-lite,uk,plaintiff_loosing_share,0.0,0.6889,0,0.0,0.6889,0
+gemini/gemini-3.1-flash-lite,uk,plaintiff_no1_ISIC1_industry_category,0.0,0.8222,0,0.0,0.8222,0
+gemini/gemini-3.1-flash-lite,uk,plaintiffs_all_count,1.0,0.9444,1,1.0,0.9444,1
+gemini/gemini-3.1-flash-lite,uk,trial_end_date,0.3333,0.6667,0,0.3333,0.6667,0
+gemini/gemini-3.1-flash-lite,uk,trial_start_date,0.0,0.7333,0,0.0,0.76,0
+gemini/gemini-3.1-flash-lite,us,court_cost_awarded_nominal,1.0,1.0,1,,,
+gemini/gemini-3.1-flash-lite,us,defendant_no1_ISIC1_industry_category,0.0,0.5444,0,0.0,0.5444,0
+gemini/gemini-3.1-flash-lite,us,defendants_all_count,1.0,0.9,1,1.0,0.9,1
+gemini/gemini-3.1-flash-lite,us,dispute_value_nominal,0.3333,0.8556,0,0.3333,0.8556,0
+gemini/gemini-3.1-flash-lite,us,party_compensation_awarded_nominal,1.0,1.0,1,,,
+gemini/gemini-3.1-flash-lite,us,plaintiff_loosing_share,0.0,0.1667,0,0.0,0.1667,0
+gemini/gemini-3.1-flash-lite,us,plaintiff_no1_ISIC1_industry_category,0.0,0.5778,0,0.0,0.5778,0
+gemini/gemini-3.1-flash-lite,us,plaintiffs_all_count,1.0,0.9,1,1.0,0.9,1
+gemini/gemini-3.1-flash-lite,us,trial_end_date,1.0,1.0,1,1.0,1.0,1
+gemini/gemini-3.1-flash-lite,us,trial_start_date,0.0,0.5667,0,0.0,0.56,0

data/analysis/iaa/alt_test_reference_gpt-5.4-mini.csv

@@ -0,0 +1,81 @@
+candidate,country,field,winning_rate,advantage_probability,passes,winning_rate_nontrivial,advantage_probability_nontrivial,passes_nontrivial
+gpt-5.4-mini,br,court_cost_awarded_nominal,1.0,1.0,1,,,
+gpt-5.4-mini,br,defendant_no1_ISIC1_industry_category,0.3333,0.852,0,0.6667,0.8869,1
+gpt-5.4-mini,br,defendants_all_count,1.0,0.9556,1,1.0,0.9556,1
+gpt-5.4-mini,br,dispute_value_nominal,0.3333,0.8491,0,0.0,0.8667,0
+gpt-5.4-mini,br,party_compensation_awarded_nominal,1.0,0.9602,1,,,
+gpt-5.4-mini,br,plaintiff_loosing_share,0.3333,0.8871,0,0.3333,0.8889,0
+gpt-5.4-mini,br,plaintiff_no1_ISIC1_industry_category,0.6667,0.8632,1,0.6667,0.8654,1
+gpt-5.4-mini,br,plaintiffs_all_count,1.0,0.9778,1,1.0,0.9778,1
+gpt-5.4-mini,br,trial_end_date,1.0,0.9556,1,1.0,1.0,1
+gpt-5.4-mini,br,trial_start_date,0.0,0.6994,0,,,
+gpt-5.4-mini,ch,court_cost_awarded_nominal,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,ch,defendant_no1_ISIC1_industry_category,0.0,0.7889,0,0.3333,0.9444,0
+gpt-5.4-mini,ch,defendants_all_count,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,ch,dispute_value_nominal,0.0,0.8444,0,0.6667,0.9074,1
+gpt-5.4-mini,ch,party_compensation_awarded_nominal,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,ch,plaintiff_loosing_share,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,ch,plaintiff_no1_ISIC1_industry_category,0.0,0.5667,0,0.0,0.8125,0
+gpt-5.4-mini,ch,plaintiffs_all_count,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,ch,trial_end_date,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,ch,trial_start_date,0.0,0.7444,0,0.3333,0.8889,0
+gpt-5.4-mini,de,court_cost_awarded_nominal,1.0,1.0,1,,,
+gpt-5.4-mini,de,defendant_no1_ISIC1_industry_category,0.3333,0.8,0,0.6667,0.8528,1
+gpt-5.4-mini,de,defendants_all_count,0.0,0.8889,0,0.0,0.8889,0
+gpt-5.4-mini,de,dispute_value_nominal,1.0,0.9389,1,1.0,0.9312,1
+gpt-5.4-mini,de,party_compensation_awarded_nominal,1.0,1.0,1,,,
+gpt-5.4-mini,de,plaintiff_loosing_share,0.3333,0.7667,0,0.3333,0.7568,0
+gpt-5.4-mini,de,plaintiff_no1_ISIC1_industry_category,0.6667,0.9444,1,0.6667,0.9437,1
+gpt-5.4-mini,de,plaintiffs_all_count,0.3333,0.9333,0,0.3333,0.9333,0
+gpt-5.4-mini,de,trial_end_date,0.0,0.1333,0,0.0,0.1333,0
+gpt-5.4-mini,de,trial_start_date,0.0,0.5556,0,,,
+gpt-5.4-mini,ge,court_cost_awarded_nominal,0.6667,0.8167,1,0.6667,0.8058,1
+gpt-5.4-mini,ge,defendant_no1_ISIC1_industry_category,0.0,0.7755,0,0.6667,0.8559,1
+gpt-5.4-mini,ge,defendants_all_count,1.0,0.9326,1,1.0,0.9326,1
+gpt-5.4-mini,ge,dispute_value_nominal,0.0,0.6521,0,0.0,0.7172,0
+gpt-5.4-mini,ge,party_compensation_awarded_nominal,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,ge,plaintiff_loosing_share,1.0,0.8822,1,1.0,0.8822,1
+gpt-5.4-mini,ge,plaintiff_no1_ISIC1_industry_category,0.0,0.7017,0,0.0,0.7307,0
+gpt-5.4-mini,ge,plaintiffs_all_count,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,ge,trial_end_date,0.0,0.7933,0,0.0,0.7933,0
+gpt-5.4-mini,ge,trial_start_date,0.0,0.6347,0,0.0,0.6764,0
+gpt-5.4-mini,sg,court_cost_awarded_nominal,1.0,0.9506,1,1.0,1.0,1
+gpt-5.4-mini,sg,defendant_no1_ISIC1_industry_category,1.0,0.8636,1,1.0,0.8859,1
+gpt-5.4-mini,sg,defendants_all_count,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,sg,dispute_value_nominal,0.0,0.7358,0,0.0,0.8472,0
+gpt-5.4-mini,sg,party_compensation_awarded_nominal,1.0,0.9506,1,0.3333,0.9259,0
+gpt-5.4-mini,sg,plaintiff_loosing_share,0.0,0.7228,0,0.0,0.7228,0
+gpt-5.4-mini,sg,plaintiff_no1_ISIC1_industry_category,0.0,0.7395,0,0.0,0.7713,0
+gpt-5.4-mini,sg,plaintiffs_all_count,1.0,0.9586,1,1.0,0.9586,1
+gpt-5.4-mini,sg,trial_end_date,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,sg,trial_start_date,0.0,0.6611,0,0.0,0.6611,0
+gpt-5.4-mini,tw,court_cost_awarded_nominal,0.6667,0.8889,1,0.0,0.7059,0
+gpt-5.4-mini,tw,defendant_no1_ISIC1_industry_category,0.3333,0.8389,0,0.6667,0.8724,1
+gpt-5.4-mini,tw,defendants_all_count,0.3333,0.8333,0,0.3333,0.8333,0
+gpt-5.4-mini,tw,dispute_value_nominal,0.0,0.6389,0,0.0,0.8182,0
+gpt-5.4-mini,tw,party_compensation_awarded_nominal,1.0,1.0,1,,,
+gpt-5.4-mini,tw,plaintiff_loosing_share,0.0,0.8444,0,0.0,0.8141,0
+gpt-5.4-mini,tw,plaintiff_no1_ISIC1_industry_category,0.6667,0.8944,1,0.6667,0.9304,1
+gpt-5.4-mini,tw,plaintiffs_all_count,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,tw,trial_end_date,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,tw,trial_start_date,0.0,0.6222,0,0.0,0.6267,0
+gpt-5.4-mini,uk,court_cost_awarded_nominal,0.0,0.9,0,,,
+gpt-5.4-mini,uk,defendant_no1_ISIC1_industry_category,1.0,0.8889,1,1.0,0.8889,1
+gpt-5.4-mini,uk,defendants_all_count,0.0,0.8778,0,0.0,0.8778,0
+gpt-5.4-mini,uk,dispute_value_nominal,0.0,0.8111,0,0.3333,0.8472,0
+gpt-5.4-mini,uk,party_compensation_awarded_nominal,1.0,0.9333,1,,,
+gpt-5.4-mini,uk,plaintiff_loosing_share,0.3333,0.8111,0,0.3333,0.8111,0
+gpt-5.4-mini,uk,plaintiff_no1_ISIC1_industry_category,0.0,0.8,0,0.0,0.8,0
+gpt-5.4-mini,uk,plaintiffs_all_count,1.0,0.9667,1,1.0,0.9667,1
+gpt-5.4-mini,uk,trial_end_date,0.3333,0.6667,0,0.3333,0.6667,0
+gpt-5.4-mini,uk,trial_start_date,0.0,0.4667,0,0.0,0.52,0
+gpt-5.4-mini,us,court_cost_awarded_nominal,1.0,1.0,1,,,
+gpt-5.4-mini,us,defendant_no1_ISIC1_industry_category,0.0,0.6111,0,0.0,0.6111,0
+gpt-5.4-mini,us,defendants_all_count,1.0,0.9333,1,1.0,0.9333,1
+gpt-5.4-mini,us,dispute_value_nominal,0.6667,0.9111,1,0.6667,0.9111,1
+gpt-5.4-mini,us,party_compensation_awarded_nominal,1.0,1.0,1,,,
+gpt-5.4-mini,us,plaintiff_loosing_share,0.0,0.5,0,0.0,0.5,0
+gpt-5.4-mini,us,plaintiff_no1_ISIC1_industry_category,0.0,0.5778,0,0.0,0.5778,0
+gpt-5.4-mini,us,plaintiffs_all_count,0.0,0.8333,0,0.0,0.8333,0
+gpt-5.4-mini,us,trial_end_date,1.0,1.0,1,1.0,1.0,1
+gpt-5.4-mini,us,trial_start_date,0.0,0.4333,0,0.0,0.48,0

data/analysis/iaa/alt_test_reference_harvey.csv

@@ -0,0 +1,81 @@
+candidate,country,field,winning_rate,advantage_probability,passes,winning_rate_nontrivial,advantage_probability_nontrivial,passes_nontrivial
+harvey,br,court_cost_awarded_nominal,1.0,1.0,1,,,
+harvey,br,defendant_no1_ISIC1_industry_category,0.3333,0.8298,0,0.3333,0.8591,0
+harvey,br,defendants_all_count,0.6667,0.8538,1,0.6667,0.8538,1
+harvey,br,dispute_value_nominal,1.0,1.0,1,1.0,1.0,1
+harvey,br,party_compensation_awarded_nominal,1.0,0.9778,1,,,
+harvey,br,plaintiff_loosing_share,0.0,0.7743,0,0.0,0.7222,0
+harvey,br,plaintiff_no1_ISIC1_industry_category,0.6667,0.8632,1,0.6667,0.8397,1
+harvey,br,plaintiffs_all_count,0.0,0.8316,0,0.0,0.8316,0
+harvey,br,trial_end_date,1.0,1.0,1,1.0,1.0,1
+harvey,br,trial_start_date,1.0,0.9778,1,,,
+harvey,ch,court_cost_awarded_nominal,1.0,0.9333,1,1.0,0.9333,1
+harvey,ch,defendant_no1_ISIC1_industry_category,1.0,0.9556,1,1.0,0.9444,1
+harvey,ch,defendants_all_count,1.0,0.9667,1,1.0,0.9667,1
+harvey,ch,dispute_value_nominal,0.3333,0.8889,0,0.3333,0.8704,0
+harvey,ch,party_compensation_awarded_nominal,1.0,0.9778,1,1.0,0.9649,1
+harvey,ch,plaintiff_loosing_share,0.0,0.7889,0,0.0,0.7889,0
+harvey,ch,plaintiff_no1_ISIC1_industry_category,1.0,0.9111,1,0.3333,0.8958,0
+harvey,ch,plaintiffs_all_count,1.0,0.9333,1,1.0,0.9333,1
+harvey,ch,trial_end_date,1.0,1.0,1,1.0,1.0,1
+harvey,ch,trial_start_date,1.0,0.9667,1,1.0,0.9583,1
+harvey,de,court_cost_awarded_nominal,1.0,0.9611,1,,,
+harvey,de,defendant_no1_ISIC1_industry_category,0.3333,0.7778,0,0.6667,0.8289,1
+harvey,de,defendants_all_count,0.0,0.8778,0,0.0,0.8778,0
+harvey,de,dispute_value_nominal,1.0,0.9389,1,1.0,0.9568,1
+harvey,de,party_compensation_awarded_nominal,1.0,0.9611,1,,,
+harvey,de,plaintiff_loosing_share,0.0,0.6944,0,0.0,0.6818,0
+harvey,de,plaintiff_no1_ISIC1_industry_category,1.0,1.0,1,1.0,1.0,1
+harvey,de,plaintiffs_all_count,1.0,0.9556,1,1.0,0.9556,1
+harvey,de,trial_end_date,0.0,0.1056,0,0.0,0.1056,0
+harvey,de,trial_start_date,1.0,1.0,1,,,
+harvey,ge,court_cost_awarded_nominal,0.6667,0.8889,1,0.6667,0.8822,1
+harvey,ge,defendant_no1_ISIC1_industry_category,1.0,0.91,1,0.6667,0.9213,1
+harvey,ge,defendants_all_count,0.0,0.8989,0,0.0,0.8989,0
+harvey,ge,dispute_value_nominal,0.0,0.5902,0,0.0,0.6461,0
+harvey,ge,party_compensation_awarded_nominal,1.0,1.0,1,1.0,1.0,1
+harvey,ge,plaintiff_loosing_share,0.6667,0.8544,1,0.6667,0.8544,1
+harvey,ge,plaintiff_no1_ISIC1_industry_category,1.0,0.8874,1,0.3333,0.8828,0
+harvey,ge,plaintiffs_all_count,0.0,0.8822,0,0.0,0.8822,0
+harvey,ge,trial_end_date,0.0,0.5395,0,0.0,0.5395,0
+harvey,ge,trial_start_date,0.6667,0.904,1,0.3333,0.8836,0
+harvey,sg,court_cost_awarded_nominal,1.0,1.0,1,1.0,1.0,1
+harvey,sg,defendant_no1_ISIC1_industry_category,1.0,0.942,1,1.0,0.941,1
+harvey,sg,defendants_all_count,1.0,1.0,1,1.0,1.0,1
+harvey,sg,dispute_value_nominal,0.6667,0.8926,1,0.0,0.8843,0
+harvey,sg,party_compensation_awarded_nominal,1.0,0.9506,1,0.3333,0.9259,0
+harvey,sg,plaintiff_loosing_share,0.0,0.7228,0,0.0,0.7228,0
+harvey,sg,plaintiff_no1_ISIC1_industry_category,1.0,0.9463,1,1.0,0.944,1
+harvey,sg,plaintiffs_all_count,1.0,1.0,1,1.0,1.0,1
+harvey,sg,trial_end_date,1.0,1.0,1,1.0,1.0,1
+harvey,sg,trial_start_date,0.3333,0.7728,0,0.3333,0.7728,0
+harvey,tw,court_cost_awarded_nominal,0.6667,0.8889,1,0.0,0.7059,0
+harvey,tw,defendant_no1_ISIC1_industry_category,0.6667,0.9222,1,0.6667,0.9189,1
+harvey,tw,defendants_all_count,0.0,0.8333,0,0.0,0.8333,0
+harvey,tw,dispute_value_nominal,0.0,0.8389,0,0.0,0.9091,0
+harvey,tw,party_compensation_awarded_nominal,1.0,1.0,1,,,
+harvey,tw,plaintiff_loosing_share,0.0,0.6833,0,0.0,0.6234,0
+harvey,tw,plaintiff_no1_ISIC1_industry_category,1.0,0.9778,1,1.0,0.977,1
+harvey,tw,plaintiffs_all_count,1.0,0.9611,1,1.0,0.9611,1
+harvey,tw,trial_end_date,1.0,1.0,1,1.0,1.0,1
+harvey,tw,trial_start_date,0.0,0.3556,0,0.0,0.2,0
+harvey,uk,court_cost_awarded_nominal,1.0,1.0,1,,,
+harvey,uk,defendant_no1_ISIC1_industry_category,1.0,0.9444,1,1.0,0.9444,1
+harvey,uk,defendants_all_count,0.0,0.8778,0,0.0,0.8778,0
+harvey,uk,dispute_value_nominal,0.3333,0.8333,0,0.3333,0.875,0
+harvey,uk,party_compensation_awarded_nominal,1.0,1.0,1,,,
+harvey,uk,plaintiff_loosing_share,0.3333,0.8,0,0.3333,0.8,0
+harvey,uk,plaintiff_no1_ISIC1_industry_category,1.0,0.9222,1,1.0,0.9222,1
+harvey,uk,plaintiffs_all_count,1.0,0.9778,1,1.0,0.9778,1
+harvey,uk,trial_end_date,1.0,1.0,1,1.0,1.0,1
+harvey,uk,trial_start_date,1.0,0.8889,1,1.0,0.9467,1
+harvey,us,court_cost_awarded_nominal,1.0,1.0,1,,,
+harvey,us,defendant_no1_ISIC1_industry_category,0.0,0.8111,0,0.0,0.8111,0
+harvey,us,defendants_all_count,1.0,0.9333,1,1.0,0.9333,1
+harvey,us,dispute_value_nominal,1.0,0.9444,1,1.0,0.9444,1
+harvey,us,party_compensation_awarded_nominal,1.0,1.0,1,,,
+harvey,us,plaintiff_loosing_share,0.0,0.4667,0,0.0,0.4667,0
+harvey,us,plaintiff_no1_ISIC1_industry_category,0.0,0.7111,0,0.0,0.7111,0
+harvey,us,plaintiffs_all_count,1.0,0.8667,1,1.0,0.8667,1
+harvey,us,trial_end_date,1.0,1.0,1,1.0,1.0,1
+harvey,us,trial_start_date,0.0,0.1,0,0.0,0.04,0

data/analysis/iaa/alt_test_reference_legora-1.csv

@@ -0,0 +1,81 @@
+candidate,country,field,winning_rate,advantage_probability,passes,winning_rate_nontrivial,advantage_probability_nontrivial,passes_nontrivial
+legora-1,br,court_cost_awarded_nominal,1.0,1.0,1,,,
+legora-1,br,defendant_no1_ISIC1_industry_category,1.0,0.9285,1,1.0,0.9623,1
+legora-1,br,defendants_all_count,0.0,0.7765,0,0.0,0.7765,0
+legora-1,br,dispute_value_nominal,1.0,1.0,1,1.0,1.0,1
+legora-1,br,party_compensation_awarded_nominal,1.0,1.0,1,,,
+legora-1,br,plaintiff_loosing_share,0.0,0.689,0,0.0,0.6135,0
+legora-1,br,plaintiff_no1_ISIC1_industry_category,1.0,1.0,1,1.0,1.0,1
+legora-1,br,plaintiffs_all_count,0.0,0.772,0,0.0,0.772,0
+legora-1,br,trial_end_date,1.0,0.977,1,1.0,0.971,1
+legora-1,br,trial_start_date,1.0,1.0,1,,,
+legora-1,ch,court_cost_awarded_nominal,1.0,1.0,1,1.0,1.0,1
+legora-1,ch,defendant_no1_ISIC1_industry_category,1.0,1.0,1,1.0,1.0,1
+legora-1,ch,defendants_all_count,1.0,1.0,1,1.0,1.0,1
+legora-1,ch,dispute_value_nominal,0.6667,0.9111,1,0.3333,0.8519,0
+legora-1,ch,party_compensation_awarded_nominal,1.0,0.9556,1,0.3333,0.9298,0
+legora-1,ch,plaintiff_loosing_share,1.0,0.9667,1,1.0,0.9667,1
+legora-1,ch,plaintiff_no1_ISIC1_industry_category,1.0,1.0,1,1.0,1.0,1
+legora-1,ch,plaintiffs_all_count,1.0,1.0,1,1.0,1.0,1
+legora-1,ch,trial_end_date,1.0,1.0,1,1.0,1.0,1
+legora-1,ch,trial_start_date,0.3333,0.9111,0,0.3333,0.8889,0
+legora-1,de,court_cost_awarded_nominal,1.0,1.0,1,,,
+legora-1,de,defendant_no1_ISIC1_industry_category,1.0,0.9189,1,1.0,0.9383,1
+legora-1,de,defendants_all_count,0.0,0.3176,0,0.0,0.3176,0
+legora-1,de,dispute_value_nominal,0.3333,0.8784,0,0.0,0.8644,0
+legora-1,de,party_compensation_awarded_nominal,1.0,1.0,1,,,
+legora-1,de,plaintiff_loosing_share,0.0,0.6473,0,0.0,0.6323,0
+legora-1,de,plaintiff_no1_ISIC1_industry_category,1.0,0.971,1,1.0,0.9706,1
+legora-1,de,plaintiffs_all_count,0.0,0.1621,0,0.0,0.1621,0
+legora-1,de,trial_end_date,0.0,0.098,0,0.0,0.098,0
+legora-1,de,trial_start_date,1.0,1.0,1,,,
+legora-1,ge,court_cost_awarded_nominal,,,,,,
+legora-1,ge,defendant_no1_ISIC1_industry_category,,,,,,
+legora-1,ge,defendants_all_count,,,,,,
+legora-1,ge,dispute_value_nominal,,,,,,
+legora-1,ge,party_compensation_awarded_nominal,,,,,,
+legora-1,ge,plaintiff_loosing_share,,,,,,
+legora-1,ge,plaintiff_no1_ISIC1_industry_category,,,,,,
+legora-1,ge,plaintiffs_all_count,,,,,,
+legora-1,ge,trial_end_date,,,,,,
+legora-1,ge,trial_start_date,,,,,,
+legora-1,sg,court_cost_awarded_nominal,1.0,1.0,1,1.0,1.0,1
+legora-1,sg,defendant_no1_ISIC1_industry_category,1.0,0.9753,1,1.0,0.9744,1
+legora-1,sg,defendants_all_count,0.6667,0.9296,1,0.6667,0.9296,1
+legora-1,sg,dispute_value_nominal,0.0,0.8389,0,0.0,0.7708,0
+legora-1,sg,party_compensation_awarded_nominal,1.0,0.9506,1,0.3333,0.9259,0
+legora-1,sg,plaintiff_loosing_share,0.0,0.4253,0,0.0,0.4253,0
+legora-1,sg,plaintiff_no1_ISIC1_industry_category,0.3333,0.8346,0,0.3333,0.8273,0
+legora-1,sg,plaintiffs_all_count,0.0,0.7272,0,0.0,0.7272,0
+legora-1,sg,trial_end_date,1.0,1.0,1,1.0,1.0,1
+legora-1,sg,trial_start_date,0.3333,0.7728,0,0.3333,0.7728,0
+legora-1,tw,court_cost_awarded_nominal,0.6667,0.8889,1,0.0,0.7059,0
+legora-1,tw,defendant_no1_ISIC1_industry_category,1.0,0.9778,1,1.0,0.977,1
+legora-1,tw,defendants_all_count,0.3333,0.8944,0,0.3333,0.8944,0
+legora-1,tw,dispute_value_nominal,0.0,0.8444,0,0.0,0.8182,0
+legora-1,tw,party_compensation_awarded_nominal,1.0,1.0,1,,,
+legora-1,tw,plaintiff_loosing_share,0.0,0.8444,0,0.0,0.8141,0
+legora-1,tw,plaintiff_no1_ISIC1_industry_category,1.0,0.9778,1,1.0,0.977,1
+legora-1,tw,plaintiffs_all_count,1.0,1.0,1,1.0,1.0,1
+legora-1,tw,trial_end_date,1.0,1.0,1,1.0,1.0,1
+legora-1,tw,trial_start_date,0.0,0.3556,0,0.0,0.2,0
+legora-1,uk,court_cost_awarded_nominal,1.0,1.0,1,,,
+legora-1,uk,defendant_no1_ISIC1_industry_category,1.0,0.9111,1,1.0,0.9111,1
+legora-1,uk,defendants_all_count,0.6667,0.9222,1,0.6667,0.9222,1
+legora-1,uk,dispute_value_nominal,1.0,0.8778,1,0.0,0.8472,0
+legora-1,uk,party_compensation_awarded_nominal,1.0,0.9667,1,,,
+legora-1,uk,plaintiff_loosing_share,0.0,0.6222,0,0.0,0.6222,0
+legora-1,uk,plaintiff_no1_ISIC1_industry_category,0.0,0.8111,0,0.0,0.8111,0
+legora-1,uk,plaintiffs_all_count,0.6667,0.9222,1,0.6667,0.9222,1
+legora-1,uk,trial_end_date,1.0,1.0,1,1.0,1.0,1
+legora-1,uk,trial_start_date,1.0,0.9333,1,1.0,0.92,1
+legora-1,us,court_cost_awarded_nominal,1.0,1.0,1,,,
+legora-1,us,defendant_no1_ISIC1_industry_category,0.0,0.7333,0,0.0,0.7333,0
+legora-1,us,defendants_all_count,1.0,0.9333,1,1.0,0.9333,1
+legora-1,us,dispute_value_nominal,0.0,0.5111,0,0.0,0.5111,0
+legora-1,us,party_compensation_awarded_nominal,1.0,1.0,1,,,
+legora-1,us,plaintiff_loosing_share,0.0,0.1,0,0.0,0.1,0
+legora-1,us,plaintiff_no1_ISIC1_industry_category,0.0,0.7333,0,0.0,0.7333,0
+legora-1,us,plaintiffs_all_count,0.6667,0.8667,1,0.6667,0.8667,1
+legora-1,us,trial_end_date,1.0,1.0,1,1.0,1.0,1
+legora-1,us,trial_start_date,0.0,0.1333,0,0.0,0.0,0

data/analysis/iaa/alt_test_reference_legora-2.csv

@@ -0,0 +1,81 @@
+candidate,country,field,winning_rate,advantage_probability,passes,winning_rate_nontrivial,advantage_probability_nontrivial,passes_nontrivial
+legora-2,br,court_cost_awarded_nominal,1.0,1.0,1,,,
+legora-2,br,defendant_no1_ISIC1_industry_category,0.6667,0.8936,1,0.6667,0.9246,1
+legora-2,br,defendants_all_count,0.0,0.8102,0,0.0,0.8102,0
+legora-2,br,dispute_value_nominal,1.0,0.9762,1,0.5,0.9231,1
+legora-2,br,party_compensation_awarded_nominal,1.0,1.0,1,,,
+legora-2,br,plaintiff_loosing_share,0.0,0.7311,0,0.0,0.6792,0
+legora-2,br,plaintiff_no1_ISIC1_industry_category,1.0,1.0,1,1.0,1.0,1
+legora-2,br,plaintiffs_all_count,0.0,0.7626,0,0.0,0.7626,0
+legora-2,br,trial_end_date,1.0,0.9762,1,1.0,0.9697,1
+legora-2,br,trial_start_date,1.0,1.0,1,,,
+legora-2,ch,court_cost_awarded_nominal,1.0,1.0,1,1.0,1.0,1
+legora-2,ch,defendant_no1_ISIC1_industry_category,0.0,0.8444,0,0.0,0.8056,0
+legora-2,ch,defendants_all_count,1.0,0.9778,1,1.0,0.9778,1
+legora-2,ch,dispute_value_nominal,1.0,0.9444,1,0.6667,0.9074,1
+legora-2,ch,party_compensation_awarded_nominal,1.0,1.0,1,1.0,1.0,1
+legora-2,ch,plaintiff_loosing_share,0.3333,0.9111,0,0.3333,0.9111,0
+legora-2,ch,plaintiff_no1_ISIC1_industry_category,1.0,1.0,1,1.0,1.0,1
+legora-2,ch,plaintiffs_all_count,1.0,0.9667,1,1.0,0.9667,1
+legora-2,ch,trial_end_date,1.0,1.0,1,1.0,1.0,1
+legora-2,ch,trial_start_date,1.0,0.9778,1,1.0,0.9722,1
+legora-2,de,court_cost_awarded_nominal,1.0,1.0,1,,,
+legora-2,de,defendant_no1_ISIC1_industry_category,1.0,0.9222,1,1.0,0.9411,1
+legora-2,de,defendants_all_count,0.0,0.1889,0,0.0,0.1889,0
+legora-2,de,dispute_value_nominal,0.6667,0.8833,1,0.0,0.8704,0
+legora-2,de,party_compensation_awarded_nominal,1.0,1.0,1,,,
+legora-2,de,plaintiff_loosing_share,0.0,0.6056,0,0.0,0.5892,0
+legora-2,de,plaintiff_no1_ISIC1_industry_category,1.0,0.9722,1,1.0,0.9718,1
+legora-2,de,plaintiffs_all_count,0.0,0.1944,0,0.0,0.1944,0
+legora-2,de,trial_end_date,0.0,0.0944,0,0.0,0.0944,0
+legora-2,de,trial_start_date,1.0,1.0,1,,,
+legora-2,ge,court_cost_awarded_nominal,,,,,,
+legora-2,ge,defendant_no1_ISIC1_industry_category,,,,,,
+legora-2,ge,defendants_all_count,,,,,,
+legora-2,ge,dispute_value_nominal,,,,,,
+legora-2,ge,party_compensation_awarded_nominal,,,,,,
+legora-2,ge,plaintiff_loosing_share,,,,,,
+legora-2,ge,plaintiff_no1_ISIC1_industry_category,,,,,,
+legora-2,ge,plaintiffs_all_count,,,,,,
+legora-2,ge,trial_end_date,,,,,,
+legora-2,ge,trial_start_date,,,,,,
+legora-2,sg,court_cost_awarded_nominal,1.0,1.0,1,1.0,1.0,1
+legora-2,sg,defendant_no1_ISIC1_industry_category,1.0,0.8469,1,1.0,0.8436,1
+legora-2,sg,defendants_all_count,1.0,0.971,1,1.0,0.971,1
+legora-2,sg,dispute_value_nominal,0.0,0.8389,0,0.0,0.7708,0
+legora-2,sg,party_compensation_awarded_nominal,1.0,0.9753,1,1.0,0.963,1
+legora-2,sg,plaintiff_loosing_share,0.0,0.4253,0,0.0,0.4253,0
+legora-2,sg,plaintiff_no1_ISIC1_industry_category,0.3333,0.8346,0,0.3333,0.8273,0
+legora-2,sg,plaintiffs_all_count,0.6667,0.9173,1,0.6667,0.9173,1
+legora-2,sg,trial_end_date,1.0,1.0,1,1.0,1.0,1
+legora-2,sg,trial_start_date,0.3333,0.7728,0,0.3333,0.7728,0
+legora-2,tw,court_cost_awarded_nominal,0.6667,0.8889,1,0.0,0.7059,0
+legora-2,tw,defendant_no1_ISIC1_industry_category,1.0,0.9778,1,1.0,0.977,1
+legora-2,tw,defendants_all_count,0.3333,0.8333,0,0.3333,0.8333,0
+legora-2,tw,dispute_value_nominal,0.0,0.9,0,1.0,1.0,1
+legora-2,tw,party_compensation_awarded_nominal,1.0,1.0,1,,,
+legora-2,tw,plaintiff_loosing_share,0.0,0.8444,0,0.0,0.8141,0
+legora-2,tw,plaintiff_no1_ISIC1_industry_category,1.0,0.9778,1,1.0,0.977,1
+legora-2,tw,plaintiffs_all_count,1.0,1.0,1,1.0,1.0,1
+legora-2,tw,trial_end_date,1.0,1.0,1,1.0,1.0,1
+legora-2,tw,trial_start_date,0.0,0.3556,0,0.0,0.2,0
+legora-2,uk,court_cost_awarded_nominal,1.0,1.0,1,,,
+legora-2,uk,defendant_no1_ISIC1_industry_category,1.0,0.9111,1,1.0,0.9111,1
+legora-2,uk,defendants_all_count,0.6667,0.9,1,0.6667,0.9,1
+legora-2,uk,dispute_value_nominal,1.0,0.9111,1,1.0,0.8889,1
+legora-2,uk,party_compensation_awarded_nominal,1.0,0.9667,1,,,
+legora-2,uk,plaintiff_loosing_share,0.0,0.6444,0,0.0,0.6444,0
+legora-2,uk,plaintiff_no1_ISIC1_industry_category,0.6667,0.8556,1,0.6667,0.8556,1
+legora-2,uk,plaintiffs_all_count,0.6667,0.9333,1,0.6667,0.9333,1
+legora-2,uk,trial_end_date,1.0,1.0,1,1.0,1.0,1
+legora-2,uk,trial_start_date,1.0,0.9556,1,1.0,0.9467,1
+legora-2,us,court_cost_awarded_nominal,1.0,1.0,1,,,
+legora-2,us,defendant_no1_ISIC1_industry_category,0.0,0.6444,0,0.0,0.6444,0
+legora-2,us,defendants_all_count,1.0,0.9667,1,1.0,0.9667,1
+legora-2,us,dispute_value_nominal,0.0,0.5778,0,0.0,0.5778,0
+legora-2,us,party_compensation_awarded_nominal,1.0,1.0,1,,,
+legora-2,us,plaintiff_loosing_share,0.0,0.1667,0,0.0,0.1667,0
+legora-2,us,plaintiff_no1_ISIC1_industry_category,0.0,0.8444,0,0.0,0.8444,0
+legora-2,us,plaintiffs_all_count,0.0,0.8,0,0.0,0.8,0
+legora-2,us,trial_end_date,1.0,1.0,1,1.0,1.0,1
+legora-2,us,trial_start_date,0.0,0.1333,0,0.0,0.0,0

data/analysis/iaa/pairwise_agreement.csv

@@ -0,0 +1,265 @@
+annotator_a,annotator_b,country,field,n,pct_exact,pct_tolerant,cohen_kappa
+93839740f5,ab2d3cfed5,br,legal_subject_judgement,19,0.0000,0.0000,0.0000
+93839740f5,ab2d3cfed5,br,trial_start_date,19,0.9474,0.9474,0.4865
+93839740f5,ab2d3cfed5,br,trial_end_date,19,0.4737,0.4737,0.3871
+93839740f5,ab2d3cfed5,br,dispute_value_nominal,19,0.5263,0.5263,0.0284
+93839740f5,ab2d3cfed5,br,plaintiff_loosing_share,19,0.8421,0.8421,0.7489
+93839740f5,ab2d3cfed5,br,court_cost_awarded_nominal,19,0.8947,0.8947,0.0000
+93839740f5,ab2d3cfed5,br,party_compensation_awarded_nominal,19,0.8421,0.8421,-0.0364
+93839740f5,ab2d3cfed5,br,plaintiffs_all_count,19,0.7368,0.7368,0.4663
+93839740f5,ab2d3cfed5,br,defendants_all_count,19,0.7368,0.7368,0.5561
+93839740f5,ab2d3cfed5,br,plaintiff_no1_ISIC1_industry_category,19,0.5789,0.5789,0.4349
+93839740f5,ab2d3cfed5,br,defendant_no1_ISIC1_industry_category,19,0.6316,0.6316,0.5181
+93839740f5,primary,br,legal_subject_judgement,19,0.0000,0.0000,0.0000
+93839740f5,primary,br,trial_start_date,19,0.7895,0.7895,0.1648
+93839740f5,primary,br,trial_end_date,19,0.4211,0.4211,0.3428
+93839740f5,primary,br,dispute_value_nominal,19,0.6316,0.6316,0.2083
+93839740f5,primary,br,plaintiff_loosing_share,19,0.3684,0.4211,0.1128
+93839740f5,primary,br,court_cost_awarded_nominal,19,0.9474,0.9474,0.0000
+93839740f5,primary,br,party_compensation_awarded_nominal,19,0.8947,0.8947,-0.0270
+93839740f5,primary,br,plaintiffs_all_count,19,0.6842,0.6842,0.1024
+93839740f5,primary,br,defendants_all_count,19,0.4211,0.4211,0.0000
+93839740f5,primary,br,plaintiff_no1_ISIC1_industry_category,19,0.3684,0.3684,-0.0654
+93839740f5,primary,br,defendant_no1_ISIC1_industry_category,19,0.6842,0.6842,0.5169
+ab2d3cfed5,primary,br,legal_subject_judgement,30,0.0000,0.0000,0.0000
+ab2d3cfed5,primary,br,trial_start_date,30,0.8000,0.8000,0.2941
+ab2d3cfed5,primary,br,trial_end_date,30,0.6000,0.6000,0.5276
+ab2d3cfed5,primary,br,dispute_value_nominal,30,0.7333,0.7333,0.3634
+ab2d3cfed5,primary,br,plaintiff_loosing_share,30,0.6000,0.6333,0.4203
+ab2d3cfed5,primary,br,court_cost_awarded_nominal,30,0.9667,1.0000,0.6512
+ab2d3cfed5,primary,br,party_compensation_awarded_nominal,30,0.9667,1.0000,0.6512
+ab2d3cfed5,primary,br,plaintiffs_all_count,30,0.7000,0.7000,0.0722
+ab2d3cfed5,primary,br,defendants_all_count,30,0.6000,0.6000,0.0909
+ab2d3cfed5,primary,br,plaintiff_no1_ISIC1_industry_category,30,0.2000,0.2000,-0.0730
+ab2d3cfed5,primary,br,defendant_no1_ISIC1_industry_category,30,0.4333,0.4333,0.2598
+5c9b11ec0c,ab2d3cfed5,ch,legal_subject_judgement,30,0.2000,0.2000,0.1946
+5c9b11ec0c,ab2d3cfed5,ch,trial_start_date,30,0.6333,0.6333,0.6005
+5c9b11ec0c,ab2d3cfed5,ch,trial_end_date,30,1.0000,1.0000,1.0000
+5c9b11ec0c,ab2d3cfed5,ch,dispute_value_nominal,30,0.5667,0.5667,0.3800
+5c9b11ec0c,ab2d3cfed5,ch,plaintiff_loosing_share,30,0.8333,0.8333,0.1379
+5c9b11ec0c,ab2d3cfed5,ch,court_cost_awarded_nominal,30,0.9333,0.9333,0.9243
+5c9b11ec0c,ab2d3cfed5,ch,party_compensation_awarded_nominal,30,0.6333,0.6333,0.4712
+5c9b11ec0c,ab2d3cfed5,ch,plaintiffs_all_count,30,1.0000,1.0000,1.0000
+5c9b11ec0c,ab2d3cfed5,ch,defendants_all_count,30,0.9000,0.9000,0.7554
+5c9b11ec0c,ab2d3cfed5,ch,plaintiff_no1_ISIC1_industry_category,30,0.5667,0.5667,0.2073
+5c9b11ec0c,ab2d3cfed5,ch,defendant_no1_ISIC1_industry_category,30,0.6333,0.6333,0.5292
+5c9b11ec0c,primary,ch,legal_subject_judgement,30,0.0333,0.0333,0.0323
+5c9b11ec0c,primary,ch,trial_start_date,30,0.6333,0.6333,0.6005
+5c9b11ec0c,primary,ch,trial_end_date,30,0.9667,0.9667,0.9654
+5c9b11ec0c,primary,ch,dispute_value_nominal,30,0.6000,0.6000,0.4098
+5c9b11ec0c,primary,ch,plaintiff_loosing_share,30,0.8333,0.8333,0.1379
+5c9b11ec0c,primary,ch,court_cost_awarded_nominal,30,0.9333,0.9333,0.9243
+5c9b11ec0c,primary,ch,party_compensation_awarded_nominal,30,0.6333,0.6333,0.4712
+5c9b11ec0c,primary,ch,plaintiffs_all_count,30,1.0000,1.0000,1.0000
+5c9b11ec0c,primary,ch,defendants_all_count,30,0.9333,0.9333,0.8365
+5c9b11ec0c,primary,ch,plaintiff_no1_ISIC1_industry_category,30,0.8000,0.8000,0.3478
+5c9b11ec0c,primary,ch,defendant_no1_ISIC1_industry_category,30,0.7333,0.7333,0.6232
+ab2d3cfed5,primary,ch,legal_subject_judgement,30,0.0667,0.0667,0.0646
+ab2d3cfed5,primary,ch,trial_start_date,30,0.9667,0.9667,0.9616
+ab2d3cfed5,primary,ch,trial_end_date,30,0.9667,0.9667,0.9654
+ab2d3cfed5,primary,ch,dispute_value_nominal,30,0.8000,0.8000,0.7196
+ab2d3cfed5,primary,ch,plaintiff_loosing_share,30,1.0000,1.0000,1.0000
+ab2d3cfed5,primary,ch,court_cost_awarded_nominal,30,1.0000,1.0000,1.0000
+ab2d3cfed5,primary,ch,party_compensation_awarded_nominal,30,1.0000,1.0000,1.0000
+ab2d3cfed5,primary,ch,plaintiffs_all_count,30,1.0000,1.0000,1.0000
+ab2d3cfed5,primary,ch,defendants_all_count,30,0.9667,0.9667,0.9229
+ab2d3cfed5,primary,ch,plaintiff_no1_ISIC1_industry_category,30,0.6333,0.6333,0.3738
+ab2d3cfed5,primary,ch,defendant_no1_ISIC1_industry_category,30,0.8000,0.8000,0.7361
+40dd94f4cc,ab2d3cfed5,de,legal_subject_judgement,20,0.0000,0.0000,0.0000
+40dd94f4cc,ab2d3cfed5,de,trial_start_date,20,1.0000,1.0000,
+40dd94f4cc,ab2d3cfed5,de,trial_end_date,20,0.9500,0.9500,0.9472
+40dd94f4cc,ab2d3cfed5,de,dispute_value_nominal,20,0.2500,0.8000,0.2105
+40dd94f4cc,ab2d3cfed5,de,plaintiff_loosing_share,20,0.9000,0.9000,0.8214
+40dd94f4cc,ab2d3cfed5,de,court_cost_awarded_nominal,20,1.0000,1.0000,1.0000
+40dd94f4cc,ab2d3cfed5,de,party_compensation_awarded_nominal,20,1.0000,1.0000,1.0000
+40dd94f4cc,ab2d3cfed5,de,plaintiffs_all_count,20,0.9000,0.9000,0.4667
+40dd94f4cc,ab2d3cfed5,de,defendants_all_count,20,0.8000,0.8000,0.2727
+40dd94f4cc,ab2d3cfed5,de,plaintiff_no1_ISIC1_industry_category,20,0.3000,0.3000,0.2802
+40dd94f4cc,ab2d3cfed5,de,defendant_no1_ISIC1_industry_category,20,0.4000,0.4000,0.3043
+40dd94f4cc,primary,de,legal_subject_judgement,20,0.0000,0.0000,0.0000
+40dd94f4cc,primary,de,trial_start_date,20,1.0000,1.0000,
+40dd94f4cc,primary,de,trial_end_date,20,0.9500,0.9500,0.9472
+40dd94f4cc,primary,de,dispute_value_nominal,20,0.2000,0.7500,0.1579
+40dd94f4cc,primary,de,plaintiff_loosing_share,20,0.6000,0.8500,0.3701
+40dd94f4cc,primary,de,court_cost_awarded_nominal,20,0.9500,0.9500,0.6491
+40dd94f4cc,primary,de,party_compensation_awarded_nominal,20,0.9500,1.0000,0.0000
+40dd94f4cc,primary,de,plaintiffs_all_count,20,0.9000,0.9000,0.5652
+40dd94f4cc,primary,de,defendants_all_count,20,0.9000,0.9000,0.5652
+40dd94f4cc,primary,de,plaintiff_no1_ISIC1_industry_category,20,0.5500,0.5500,0.2885
+40dd94f4cc,primary,de,defendant_no1_ISIC1_industry_category,20,0.4500,0.4500,0.2414
+ab2d3cfed5,primary,de,legal_subject_judgement,30,0.0000,0.0000,0.0000
+ab2d3cfed5,primary,de,trial_start_date,30,1.0000,1.0000,
+ab2d3cfed5,primary,de,trial_end_date,30,0.9000,0.9000,0.8967
+ab2d3cfed5,primary,de,dispute_value_nominal,30,0.7667,0.7667,0.7485
+ab2d3cfed5,primary,de,plaintiff_loosing_share,30,0.5667,0.8667,0.3445
+ab2d3cfed5,primary,de,court_cost_awarded_nominal,30,0.9333,0.9667,0.4737
+ab2d3cfed5,primary,de,party_compensation_awarded_nominal,30,0.9667,1.0000,0.0000
+ab2d3cfed5,primary,de,plaintiffs_all_count,30,0.9000,0.9000,0.5288
+ab2d3cfed5,primary,de,defendants_all_count,30,0.8000,0.8333,0.4690
+ab2d3cfed5,primary,de,plaintiff_no1_ISIC1_industry_category,30,0.0667,0.0667,0.0367
+ab2d3cfed5,primary,de,defendant_no1_ISIC1_industry_category,30,0.1333,0.1333,0.0299
+359f544380,dd79b903af,ge,legal_subject_judgement,19,0.0000,0.0000,0.0000
+359f544380,dd79b903af,ge,trial_start_date,19,0.5263,0.5263,-0.0491
+359f544380,dd79b903af,ge,trial_end_date,19,0.2632,0.2632,0.2486
+359f544380,dd79b903af,ge,dispute_value_nominal,19,0.3158,0.5789,0.2400
+359f544380,dd79b903af,ge,plaintiff_loosing_share,19,0.0526,0.0526,-0.0523
+359f544380,dd79b903af,ge,court_cost_awarded_nominal,19,0.0526,0.7368,0.0058
+359f544380,dd79b903af,ge,party_compensation_awarded_nominal,19,0.0526,0.8421,0.0000
+359f544380,dd79b903af,ge,plaintiffs_all_count,19,0.8947,0.8947,-0.0270
+359f544380,dd79b903af,ge,defendants_all_count,19,0.7368,0.7368,0.5433
+359f544380,dd79b903af,ge,plaintiff_no1_ISIC1_industry_category,19,0.6842,0.6842,0.5366
+359f544380,dd79b903af,ge,defendant_no1_ISIC1_industry_category,19,0.7368,0.7368,0.5701
+359f544380,primary,ge,legal_subject_judgement,20,0.0000,0.0000,0.0000
+359f544380,primary,ge,trial_start_date,20,0.3000,0.3000,0.0939
+359f544380,primary,ge,trial_end_date,20,0.8000,0.8000,0.7861
+359f544380,primary,ge,dispute_value_nominal,20,0.2000,0.4000,0.1398
+359f544380,primary,ge,plaintiff_loosing_share,20,0.5500,0.5500,0.2969
+359f544380,primary,ge,court_cost_awarded_nominal,20,0.5500,0.7000,0.4156
+359f544380,primary,ge,party_compensation_awarded_nominal,20,0.2000,0.9000,0.1601
+359f544380,primary,ge,plaintiffs_all_count,20,0.9000,0.9000,-0.0256
+359f544380,primary,ge,defendants_all_count,20,0.7000,0.7000,0.5000
+359f544380,primary,ge,plaintiff_no1_ISIC1_industry_category,20,0.2500,0.2500,0.2021
+359f544380,primary,ge,defendant_no1_ISIC1_industry_category,20,0.4000,0.4000,0.2793
+dd79b903af,primary,ge,legal_subject_judgement,29,0.0000,0.0000,0.0000
+dd79b903af,primary,ge,trial_start_date,29,0.4828,0.4828,0.4025
+dd79b903af,primary,ge,trial_end_date,29,0.4483,0.4483,0.4307
+dd79b903af,primary,ge,dispute_value_nominal,29,0.4828,0.4828,0.3006
+dd79b903af,primary,ge,plaintiff_loosing_share,29,0.3793,0.4483,0.0578
+dd79b903af,primary,ge,court_cost_awarded_nominal,29,0.1724,0.1724,0.0266
+dd79b903af,primary,ge,party_compensation_awarded_nominal,29,0.6897,0.6897,0.0000
+dd79b903af,primary,ge,plaintiffs_all_count,29,0.9655,0.9655,0.6548
+dd79b903af,primary,ge,defendants_all_count,29,0.8621,0.8621,0.7184
+dd79b903af,primary,ge,plaintiff_no1_ISIC1_industry_category,29,0.4138,0.4138,0.3830
+dd79b903af,primary,ge,defendant_no1_ISIC1_industry_category,29,0.3448,0.3448,0.2218
+c7cd8dc26a,dd79b903af,sg,legal_subject_judgement,20,0.0000,0.0000,0.0000
+c7cd8dc26a,dd79b903af,sg,trial_start_date,20,0.2500,0.2500,0.1957
+c7cd8dc26a,dd79b903af,sg,trial_end_date,20,0.9500,0.9500,0.9475
+c7cd8dc26a,dd79b903af,sg,dispute_value_nominal,20,0.5000,0.5000,0.3590
+c7cd8dc26a,dd79b903af,sg,plaintiff_loosing_share,20,0.4500,0.4500,0.2857
+c7cd8dc26a,dd79b903af,sg,court_cost_awarded_nominal,20,0.5500,0.5500,0.0000
+c7cd8dc26a,dd79b903af,sg,party_compensation_awarded_nominal,20,0.3000,0.3000,0.0036
+c7cd8dc26a,dd79b903af,sg,plaintiffs_all_count,20,1.0000,1.0000,1.0000
+c7cd8dc26a,dd79b903af,sg,defendants_all_count,20,0.9500,0.9500,0.9029
+c7cd8dc26a,dd79b903af,sg,plaintiff_no1_ISIC1_industry_category,20,0.1500,0.1500,-0.0119
+c7cd8dc26a,dd79b903af,sg,defendant_no1_ISIC1_industry_category,20,0.1500,0.1500,-0.0398
+c7cd8dc26a,primary,sg,legal_subject_judgement,27,0.0000,0.0000,0.0000
+c7cd8dc26a,primary,sg,trial_start_date,27,0.2593,0.2593,0.2319
+c7cd8dc26a,primary,sg,trial_end_date,27,1.0000,1.0000,1.0000
+c7cd8dc26a,primary,sg,dispute_value_nominal,27,0.6667,0.6667,0.4465
+c7cd8dc26a,primary,sg,plaintiff_loosing_share,27,0.4074,0.4074,0.1446
+c7cd8dc26a,primary,sg,court_cost_awarded_nominal,27,0.8519,0.8519,0.0000
+c7cd8dc26a,primary,sg,party_compensation_awarded_nominal,27,0.8519,0.8519,0.5940
+c7cd8dc26a,primary,sg,plaintiffs_all_count,27,0.9630,0.9630,0.9148
+c7cd8dc26a,primary,sg,defendants_all_count,27,0.8889,0.8889,0.7437
+c7cd8dc26a,primary,sg,plaintiff_no1_ISIC1_industry_category,27,0.4074,0.4074,0.3110
+c7cd8dc26a,primary,sg,defendant_no1_ISIC1_industry_category,27,0.2593,0.2593,0.2128
+dd79b903af,primary,sg,legal_subject_judgement,21,0.0000,0.0000,0.0000
+dd79b903af,primary,sg,trial_start_date,21,0.3810,0.3810,0.3607
+dd79b903af,primary,sg,trial_end_date,21,0.9048,0.9048,0.9005
+dd79b903af,primary,sg,dispute_value_nominal,21,0.3810,0.4286,0.2417
+dd79b903af,primary,sg,plaintiff_loosing_share,21,0.7143,0.7143,0.5368
+dd79b903af,primary,sg,court_cost_awarded_nominal,21,0.5714,0.7619,0.1085
+dd79b903af,primary,sg,party_compensation_awarded_nominal,21,0.3333,0.7143,-0.0173
+dd79b903af,primary,sg,plaintiffs_all_count,21,0.9524,0.9524,0.9095
+dd79b903af,primary,sg,defendants_all_count,21,0.8571,0.8571,0.7284
+dd79b903af,primary,sg,plaintiff_no1_ISIC1_industry_category,21,0.4286,0.4286,0.2941
+dd79b903af,primary,sg,defendant_no1_ISIC1_industry_category,21,0.3333,0.3333,0.2481
+ab2d3cfed5,cf5a561621,tw,legal_subject_judgement,20,0.1000,0.1000,0.0955
+ab2d3cfed5,cf5a561621,tw,trial_start_date,20,0.8500,0.8500,0.8276
+ab2d3cfed5,cf5a561621,tw,trial_end_date,20,0.9000,0.9000,0.8889
+ab2d3cfed5,cf5a561621,tw,dispute_value_nominal,20,0.9000,0.9000,0.7661
+ab2d3cfed5,cf5a561621,tw,plaintiff_loosing_share,20,0.9500,0.9500,0.9200
+ab2d3cfed5,cf5a561621,tw,court_cost_awarded_nominal,20,0.7500,0.7500,0.3827
+ab2d3cfed5,cf5a561621,tw,party_compensation_awarded_nominal,20,1.0000,1.0000,1.0000
+ab2d3cfed5,cf5a561621,tw,plaintiffs_all_count,20,0.8000,0.8000,0.3701
+ab2d3cfed5,cf5a561621,tw,defendants_all_count,20,0.6500,0.6500,0.3396
+ab2d3cfed5,cf5a561621,tw,plaintiff_no1_ISIC1_industry_category,20,0.2000,0.2000,0.1557
+ab2d3cfed5,cf5a561621,tw,defendant_no1_ISIC1_industry_category,20,0.3500,0.3500,0.2935
+ab2d3cfed5,primary,tw,legal_subject_judgement,30,0.0000,0.0000,0.0000
+ab2d3cfed5,primary,tw,trial_start_date,30,0.8333,0.8333,0.8175
+ab2d3cfed5,primary,tw,trial_end_date,30,0.9333,0.9333,0.9277
+ab2d3cfed5,primary,tw,dispute_value_nominal,30,0.9000,0.9000,0.8013
+ab2d3cfed5,primary,tw,plaintiff_loosing_share,30,0.9333,0.9667,0.8745
+ab2d3cfed5,primary,tw,court_cost_awarded_nominal,30,0.6000,0.6000,0.3271
+ab2d3cfed5,primary,tw,party_compensation_awarded_nominal,30,1.0000,1.0000,1.0000
+ab2d3cfed5,primary,tw,plaintiffs_all_count,30,0.9667,0.9667,0.8976
+ab2d3cfed5,primary,tw,defendants_all_count,30,0.8333,0.8667,0.6305
+ab2d3cfed5,primary,tw,plaintiff_no1_ISIC1_industry_category,30,0.1667,0.1667,0.0942
+ab2d3cfed5,primary,tw,defendant_no1_ISIC1_industry_category,30,0.2333,0.2333,0.1766
+cf5a561621,primary,tw,legal_subject_judgement,20,0.0000,0.0000,0.0000
+cf5a561621,primary,tw,trial_start_date,20,1.0000,1.0000,1.0000
+cf5a561621,primary,tw,trial_end_date,20,1.0000,1.0000,1.0000
+cf5a561621,primary,tw,dispute_value_nominal,20,0.9500,0.9500,0.8718
+cf5a561621,primary,tw,plaintiff_loosing_share,20,0.9000,0.9500,0.8400
+cf5a561621,primary,tw,court_cost_awarded_nominal,20,0.6500,0.6500,0.3519
+cf5a561621,primary,tw,party_compensation_awarded_nominal,20,1.0000,1.0000,1.0000
+cf5a561621,primary,tw,plaintiffs_all_count,20,0.8500,0.8500,0.4545
+cf5a561621,primary,tw,defendants_all_count,20,0.6500,0.6500,0.3269
+cf5a561621,primary,tw,plaintiff_no1_ISIC1_industry_category,20,0.9000,0.9000,0.7419
+cf5a561621,primary,tw,defendant_no1_ISIC1_industry_category,20,0.7500,0.7500,0.6700
+ab2d3cfed5,b36683f5be,uk,legal_subject_judgement,30,0.0000,0.0000,0.0000
+ab2d3cfed5,b36683f5be,uk,trial_start_date,30,0.2667,0.2667,-0.0577
+ab2d3cfed5,b36683f5be,uk,trial_end_date,30,0.0000,0.0000,0.0000
+ab2d3cfed5,b36683f5be,uk,dispute_value_nominal,30,0.5333,0.7000,0.4301
+ab2d3cfed5,b36683f5be,uk,plaintiff_loosing_share,30,0.3000,0.5333,0.1711
+ab2d3cfed5,b36683f5be,uk,court_cost_awarded_nominal,30,1.0000,1.0000,
+ab2d3cfed5,b36683f5be,uk,party_compensation_awarded_nominal,30,1.0000,1.0000,1.0000
+ab2d3cfed5,b36683f5be,uk,plaintiffs_all_count,30,0.7667,0.7667,0.3805
+ab2d3cfed5,b36683f5be,uk,defendants_all_count,30,0.7667,0.7667,0.4355
+ab2d3cfed5,b36683f5be,uk,plaintiff_no1_ISIC1_industry_category,30,0.4667,0.4667,0.3985
+ab2d3cfed5,b36683f5be,uk,defendant_no1_ISIC1_industry_category,30,0.5333,0.5333,0.4430
+ab2d3cfed5,primary,uk,legal_subject_judgement,30,0.0333,0.0333,0.0323
+ab2d3cfed5,primary,uk,trial_start_date,30,0.4000,0.4000,0.2673
+ab2d3cfed5,primary,uk,trial_end_date,30,0.5333,0.5333,0.5249
+ab2d3cfed5,primary,uk,dispute_value_nominal,30,0.4667,0.4667,0.2782
+ab2d3cfed5,primary,uk,plaintiff_loosing_share,30,0.6333,0.6667,0.4463
+ab2d3cfed5,primary,uk,court_cost_awarded_nominal,30,1.0000,1.0000,
+ab2d3cfed5,primary,uk,party_compensation_awarded_nominal,30,1.0000,1.0000,1.0000
+ab2d3cfed5,primary,uk,plaintiffs_all_count,30,0.8667,0.8667,0.6471
+ab2d3cfed5,primary,uk,defendants_all_count,30,0.8333,0.8333,0.5652
+ab2d3cfed5,primary,uk,plaintiff_no1_ISIC1_industry_category,30,0.5000,0.5000,0.4246
+ab2d3cfed5,primary,uk,defendant_no1_ISIC1_industry_category,30,0.6000,0.6000,0.5095
+b36683f5be,primary,uk,legal_subject_judgement,30,0.0333,0.0333,0.0301
+b36683f5be,primary,uk,trial_start_date,30,0.5000,0.5000,0.2475
+b36683f5be,primary,uk,trial_end_date,30,0.0000,0.0000,0.0000
+b36683f5be,primary,uk,dispute_value_nominal,30,0.3667,0.5333,0.1751
+b36683f5be,primary,uk,plaintiff_loosing_share,30,0.4333,0.4667,0.2365
+b36683f5be,primary,uk,court_cost_awarded_nominal,30,1.0000,1.0000,
+b36683f5be,primary,uk,party_compensation_awarded_nominal,30,1.0000,1.0000,1.0000
+b36683f5be,primary,uk,plaintiffs_all_count,30,0.7333,0.7333,0.3333
+b36683f5be,primary,uk,defendants_all_count,30,0.8000,0.8000,0.5041
+b36683f5be,primary,uk,plaintiff_no1_ISIC1_industry_category,30,0.3333,0.3333,0.2278
+b36683f5be,primary,uk,defendant_no1_ISIC1_industry_category,30,0.5000,0.5000,0.3767
+9bbad8ca9e,ab2d3cfed5,us,legal_subject_judgement,30,0.0667,0.0667,0.0625
+9bbad8ca9e,ab2d3cfed5,us,trial_start_date,30,1.0000,1.0000,1.0000
+9bbad8ca9e,ab2d3cfed5,us,trial_end_date,30,1.0000,1.0000,1.0000
+9bbad8ca9e,ab2d3cfed5,us,dispute_value_nominal,30,0.5667,0.5667,0.0000
+9bbad8ca9e,ab2d3cfed5,us,plaintiff_loosing_share,30,0.9667,0.9667,0.8980
+9bbad8ca9e,ab2d3cfed5,us,court_cost_awarded_nominal,30,1.0000,1.0000,
+9bbad8ca9e,ab2d3cfed5,us,party_compensation_awarded_nominal,30,1.0000,1.0000,
+9bbad8ca9e,ab2d3cfed5,us,plaintiffs_all_count,30,0.6333,0.6333,0.1247
+9bbad8ca9e,ab2d3cfed5,us,defendants_all_count,30,0.5667,0.5667,0.1034
+9bbad8ca9e,ab2d3cfed5,us,plaintiff_no1_ISIC1_industry_category,30,0.7333,0.7333,0.6347
+9bbad8ca9e,ab2d3cfed5,us,defendant_no1_ISIC1_industry_category,30,0.8667,0.8667,0.8065
+9bbad8ca9e,primary,us,legal_subject_judgement,30,0.0000,0.0000,0.0000
+9bbad8ca9e,primary,us,trial_start_date,30,1.0000,1.0000,1.0000
+9bbad8ca9e,primary,us,trial_end_date,30,1.0000,1.0000,1.0000
+9bbad8ca9e,primary,us,dispute_value_nominal,30,0.5333,0.5333,0.0000
+9bbad8ca9e,primary,us,plaintiff_loosing_share,30,0.9333,0.9333,0.8039
+9bbad8ca9e,primary,us,court_cost_awarded_nominal,30,1.0000,1.0000,
+9bbad8ca9e,primary,us,party_compensation_awarded_nominal,30,1.0000,1.0000,
+9bbad8ca9e,primary,us,plaintiffs_all_count,30,0.6333,0.6333,0.0984
+9bbad8ca9e,primary,us,defendants_all_count,30,0.5667,0.5667,0.0758
+9bbad8ca9e,primary,us,plaintiff_no1_ISIC1_industry_category,30,0.8667,0.8667,0.8071
+9bbad8ca9e,primary,us,defendant_no1_ISIC1_industry_category,30,0.9000,0.9000,0.8534
+ab2d3cfed5,primary,us,legal_subject_judgement,30,0.0000,0.0000,0.0000
+ab2d3cfed5,primary,us,trial_start_date,30,1.0000,1.0000,1.0000
+ab2d3cfed5,primary,us,trial_end_date,30,1.0000,1.0000,1.0000
+ab2d3cfed5,primary,us,dispute_value_nominal,30,0.8667,0.8667,0.7521
+ab2d3cfed5,primary,us,plaintiff_loosing_share,30,0.9000,0.9000,0.7059
+ab2d3cfed5,primary,us,court_cost_awarded_nominal,30,1.0000,1.0000,
+ab2d3cfed5,primary,us,party_compensation_awarded_nominal,30,1.0000,1.0000,
+ab2d3cfed5,primary,us,plaintiffs_all_count,30,0.6333,0.6333,0.4149
+ab2d3cfed5,primary,us,defendants_all_count,30,0.5667,0.5667,0.3810
+ab2d3cfed5,primary,us,plaintiff_no1_ISIC1_industry_category,30,0.6667,0.6667,0.5448
+ab2d3cfed5,primary,us,defendant_no1_ISIC1_industry_category,30,0.9000,0.9000,0.8534

data/analysis/paper_tables.tex

@@ -0,0 +1,63 @@
+% Auto-generated by scripts/paper_tables.py — do not edit by hand.
+% 19 release jurisdictions, 10 structured fields (free-text excluded).
+
+\begin{table*}[htbp]
+\caption{Extraction metrics by jurisdiction over the ten structured fields (recall on expert-filled cells and false-fill rate on expert-empty cells, each $\pm$1\,SE). $n$ lists the goldenset-filled/goldenset-empty denominators; where a system's coverage differs its denominators follow in parentheses (G = Gemini, H = Harvey, L = Legora); \mbox{---} marks jurisdictions a system did not cover.}
+\label{tab:metrics-by-jurisdiction}
+\centering\small
+\resizebox{\textwidth}{!}{%
+\begin{tabular}{@{}lrr@{\hskip 8pt}rr@{\hskip 8pt}rr@{\hskip 8pt}rr@{\hskip 8pt}r@{}}
+\toprule
+& \multicolumn{2}{c}{\textbf{Gemini}} & \multicolumn{2}{c}{\textbf{GPT-5.4-mini}} & \multicolumn{2}{c}{\textbf{Harvey}} & \multicolumn{2}{c}{\textbf{Legora}} & \\
+\cmidrule(lr){2-3}\cmidrule(lr){4-5}\cmidrule(lr){6-7}\cmidrule(lr){8-9}
+\textbf{Jurisdiction} & \textbf{Recall} & \textbf{False-fill} & \textbf{Recall} & \textbf{False-fill} & \textbf{Recall} & \textbf{False-fill} & \textbf{Recall} & \textbf{False-fill} & \textbf{$n$ (filled/empty)} \\
+\midrule
+Armenia & 55.1\% $\pm$ 2.3\% & 10.6\% $\pm$ 2.8\% & 58.9\% $\pm$ 2.3\% & 15.4\% $\pm$ 3.3\% & 9.4\% $\pm$ 1.4\% & 0.0\% & --- & --- & 457/123 (H: 448/122) \\
+Australia & 77.3\% $\pm$ 3.3\% & 34.3\% $\pm$ 4.1\% & 74.8\% $\pm$ 3.4\% & 50.4\% $\pm$ 4.3\% & 79.8\% $\pm$ 3.1\% & 21.2\% $\pm$ 3.5\% & 69.9\% $\pm$ 3.6\% & 5.1\% $\pm$ 1.9\% & 163/137 \\
+Belgium & 78.0\% $\pm$ 2.2\% & 9.7\% $\pm$ 2.1\% & 73.2\% $\pm$ 2.3\% & 27.2\% $\pm$ 3.2\% & 71.3\% $\pm$ 2.4\% & 8.2\% $\pm$ 2.0\% & 65.9\% $\pm$ 2.5\% & 2.6\% $\pm$ 1.1\% & 355/195 \\
+Brazil & 64.8\% $\pm$ 2.0\% & 28.0\% $\pm$ 1.7\% & 57.0\% $\pm$ 2.1\% & 36.0\% $\pm$ 1.8\% & 49.7\% $\pm$ 2.1\% & 22.0\% $\pm$ 1.5\% & 48.7\% $\pm$ 2.2\% & 11.7\% $\pm$ 1.2\% & 563/737 (L: 538/702) \\
+France & 69.1\% $\pm$ 3.3\% & 36.8\% $\pm$ 4.7\% & 78.4\% $\pm$ 3.0\% & 61.3\% $\pm$ 4.7\% & 67.5\% $\pm$ 3.4\% & 29.2\% $\pm$ 4.4\% & 64.9\% $\pm$ 3.4\% & 25.5\% $\pm$ 4.2\% & 194/106 \\
+Georgia & 55.1\% $\pm$ 1.7\% & 25.9\% $\pm$ 2.9\% & 55.2\% $\pm$ 1.7\% & 49.6\% $\pm$ 3.3\% & 40.1\% $\pm$ 1.6\% & 16.1\% $\pm$ 2.5\% & --- & --- & 896/224 \\
+Germany & 56.7\% $\pm$ 2.0\% & 29.0\% $\pm$ 1.8\% & 53.9\% $\pm$ 2.0\% & 33.0\% $\pm$ 1.8\% & 53.7\% $\pm$ 2.0\% & 16.1\% $\pm$ 1.5\% & 22.7\% $\pm$ 1.7\% & 8.4\% $\pm$ 1.2\% & 642/658 (H: 631/639) (L: 599/581) \\
+Hong Kong & 74.2\% $\pm$ 5.6\% & 31.6\% $\pm$ 7.5\% & 71.0\% $\pm$ 5.8\% & 28.9\% $\pm$ 7.4\% & 61.3\% $\pm$ 6.2\% & 10.5\% $\pm$ 5.0\% & 62.9\% $\pm$ 6.1\% & 2.6\% $\pm$ 2.6\% & 62/38 \\
+India & 73.9\% $\pm$ 3.8\% & 37.5\% $\pm$ 4.9\% & 81.3\% $\pm$ 3.4\% & 59.4\% $\pm$ 5.0\% & 78.7\% $\pm$ 3.4\% & 20.2\% $\pm$ 4.0\% & 78.0\% $\pm$ 3.5\% & 10.1\% $\pm$ 3.0\% & 134/96 (H: 141/99) (L: 141/99) \\
+Nepal & 56.8\% $\pm$ 1.6\% & 64.1\% $\pm$ 2.5\% & 59.0\% $\pm$ 1.6\% & 78.0\% $\pm$ 2.2\% & 38.8\% $\pm$ 2.0\% & 26.5\% $\pm$ 3.1\% & 34.2\% $\pm$ 3.3\% & 17.9\% $\pm$ 4.3\% & 932/368 (H: 616/204) (L: 202/78) \\
+New Zealand & 91.1\% $\pm$ 2.3\% & 32.6\% $\pm$ 4.1\% & 86.1\% $\pm$ 2.8\% & 51.5\% $\pm$ 4.3\% & 89.2\% $\pm$ 2.5\% & 14.4\% $\pm$ 3.1\% & 86.1\% $\pm$ 2.8\% & 6.1\% $\pm$ 2.1\% & 158/132 \\
+Philippines & 41.8\% $\pm$ 6.7\% & 42.2\% $\pm$ 7.4\% & 56.4\% $\pm$ 6.7\% & 48.9\% $\pm$ 7.5\% & 72.2\% $\pm$ 6.1\% & 27.8\% $\pm$ 7.5\% & 78.2\% $\pm$ 5.6\% & 24.4\% $\pm$ 6.4\% & 55/45 (H: 54/36) \\
+Serbia & 85.6\% $\pm$ 2.8\% & 40.0\% $\pm$ 4.9\% & 85.0\% $\pm$ 2.8\% & 36.0\% $\pm$ 4.8\% & 85.1\% $\pm$ 2.9\% & 10.4\% $\pm$ 3.1\% & 78.1\% $\pm$ 3.3\% & 12.0\% $\pm$ 3.2\% & 160/100 (H: 154/96) \\
+Singapore & 58.8\% $\pm$ 1.7\% & 24.2\% $\pm$ 2.2\% & 58.7\% $\pm$ 1.7\% & 30.1\% $\pm$ 2.4\% & 57.9\% $\pm$ 1.7\% & 12.0\% $\pm$ 1.7\% & 47.5\% $\pm$ 1.7\% & 5.1\% $\pm$ 1.1\% & 855/375 (G: 855/385) \\
+Spain & 75.7\% $\pm$ 1.3\% & 29.0\% $\pm$ 3.0\% & 77.7\% $\pm$ 1.3\% & 44.2\% $\pm$ 3.3\% & 53.1\% $\pm$ 4.4\% & 22.7\% $\pm$ 8.9\% & 58.6\% $\pm$ 1.5\% & 10.8\% $\pm$ 2.0\% & 1069/231 (H: 128/22) \\
+Switzerland & 88.5\% $\pm$ 1.0\% & 42.8\% $\pm$ 2.8\% & 92.0\% $\pm$ 0.9\% & 79.4\% $\pm$ 2.3\% & 87.3\% $\pm$ 1.1\% & 16.1\% $\pm$ 2.1\% & 89.5\% $\pm$ 1.0\% & 7.7\% $\pm$ 1.5\% & 989/311 \\
+Taiwan & 69.3\% $\pm$ 1.6\% & 29.5\% $\pm$ 2.1\% & 74.0\% $\pm$ 1.5\% & 43.6\% $\pm$ 2.2\% & 63.5\% $\pm$ 1.7\% & 9.8\% $\pm$ 1.3\% & 67.2\% $\pm$ 1.6\% & 4.7\% $\pm$ 1.0\% & 812/488 \\
+United Kingdom & 48.9\% $\pm$ 1.7\% & 25.2\% $\pm$ 2.0\% & 52.8\% $\pm$ 1.7\% & 28.7\% $\pm$ 2.1\% & 64.6\% $\pm$ 1.6\% & 22.0\% $\pm$ 2.0\% & 62.6\% $\pm$ 1.7\% & 9.4\% $\pm$ 1.4\% & 851/449 \\
+United States & 66.7\% $\pm$ 1.5\% & 16.2\% $\pm$ 1.9\% & 64.0\% $\pm$ 1.6\% & 22.4\% $\pm$ 2.2\% & 60.6\% $\pm$ 1.6\% & 8.9\% $\pm$ 1.5\% & 55.7\% $\pm$ 1.6\% & 7.0\% $\pm$ 1.3\% & 929/371 \\
+\bottomrule
+\end{tabular}%
+}
+\end{table*}
+
+\begin{table*}[htbp]
+\caption{Extraction metrics by field over the 19 release jurisdictions (recall and false-fill rate, each $\pm$1\,SE). $n$ as in \cref{tab:metrics-by-jurisdiction}.}
+\label{tab:metrics-by-field}
+\centering\small
+\resizebox{\textwidth}{!}{%
+\begin{tabular}{@{}lrr@{\hskip 8pt}rr@{\hskip 8pt}rr@{\hskip 8pt}rr@{\hskip 8pt}r@{}}
+\toprule
+& \multicolumn{2}{c}{\textbf{Gemini}} & \multicolumn{2}{c}{\textbf{GPT-5.4-mini}} & \multicolumn{2}{c}{\textbf{Harvey}} & \multicolumn{2}{c}{\textbf{Legora}} & \\
+\cmidrule(lr){2-3}\cmidrule(lr){4-5}\cmidrule(lr){6-7}\cmidrule(lr){8-9}
+\textbf{Variable} & \textbf{Recall} & \textbf{False-fill} & \textbf{Recall} & \textbf{False-fill} & \textbf{Recall} & \textbf{False-fill} & \textbf{Recall} & \textbf{False-fill} & \textbf{$n$ (filled/empty)} \\
+\midrule
+\texttt{court\_cost\_awarded\_nominal} & 63.9\% $\pm$ 2.1\% & 4.1\% $\pm$ 0.6\% & 65.5\% $\pm$ 2.1\% & 8.3\% $\pm$ 0.9\% & 57.8\% $\pm$ 2.4\% & 1.9\% $\pm$ 0.4\% & 70.1\% $\pm$ 2.6\% & 0.9\% $\pm$ 0.3\% & 501/1045 (G: 501/1046) (H: 427/951) (L: 308/949) \\
+\texttt{defendant\_no1\_ISIC1\_industry\_category} & 58.0\% $\pm$ 1.5\% & 65.9\% $\pm$ 2.2\% & 64.7\% $\pm$ 1.5\% & 74.7\% $\pm$ 2.0\% & 49.3\% $\pm$ 1.6\% & 37.3\% $\pm$ 2.3\% & 55.5\% $\pm$ 1.7\% & 22.5\% $\pm$ 2.1\% & 1075/471 (G: 1075/472) (H: 938/440) (L: 874/383) \\
+\texttt{defendants\_all\_count} & 76.9\% $\pm$ 1.1\% & 90.5\% $\pm$ 4.5\% & 73.8\% $\pm$ 1.1\% & 56.1\% $\pm$ 7.8\% & 71.6\% $\pm$ 1.2\% & 48.6\% $\pm$ 8.4\% & 66.6\% $\pm$ 1.3\% & 37.1\% $\pm$ 8.2\% & 1505/41 (G: 1505/42) (H: 1343/35) (L: 1222/35) \\
+\texttt{dispute\_value\_nominal} & 56.4\% $\pm$ 1.9\% & 44.0\% $\pm$ 1.7\% & 57.1\% $\pm$ 1.9\% & 53.8\% $\pm$ 1.7\% & 47.0\% $\pm$ 2.1\% & 32.9\% $\pm$ 1.7\% & 34.7\% $\pm$ 2.1\% & 12.3\% $\pm$ 1.2\% & 708/838 (G: 708/839) (H: 591/787) (L: 539/718) \\
+\texttt{party\_compensation\_awarded\_nominal} & 71.4\% $\pm$ 2.4\% & 6.6\% $\pm$ 0.7\% & 67.3\% $\pm$ 2.5\% & 13.6\% $\pm$ 1.0\% & 66.9\% $\pm$ 2.8\% & 4.7\% $\pm$ 0.6\% & 79.1\% $\pm$ 2.7\% & 4.2\% $\pm$ 0.6\% & 364/1182 (G: 364/1183) (H: 284/1094) (L: 235/1022) \\
+\texttt{plaintiff\_loosing\_share} & 71.3\% $\pm$ 1.2\% & 36.6\% $\pm$ 3.2\% & 69.9\% $\pm$ 1.3\% & 55.2\% $\pm$ 3.3\% & 71.4\% $\pm$ 1.3\% & 33.3\% $\pm$ 3.4\% & 69.3\% $\pm$ 1.4\% & 11.4\% $\pm$ 2.5\% & 1323/223 (G: 1323/224) (H: 1189/189) (L: 1090/167) \\
+\texttt{plaintiff\_no1\_ISIC1\_industry\_category} & 54.1\% $\pm$ 1.7\% & 79.5\% $\pm$ 1.5\% & 53.9\% $\pm$ 1.7\% & 80.5\% $\pm$ 1.5\% & 41.8\% $\pm$ 1.8\% & 25.4\% $\pm$ 1.7\% & 36.8\% $\pm$ 1.8\% & 16.9\% $\pm$ 1.6\% & 865/681 (G: 865/682) (H: 752/626) (L: 682/575) \\
+\texttt{plaintiffs\_all\_count} & 88.8\% $\pm$ 0.8\% & 100.0\% & 85.4\% $\pm$ 0.9\% & 71.4\% $\pm$ 12.1\% & 74.9\% $\pm$ 1.2\% & 83.3\% $\pm$ 10.8\% & 70.5\% $\pm$ 1.3\% & 28.6\% $\pm$ 12.1\% & 1532/14 (G: 1532/15) (H: 1366/12) (L: 1243/14) \\
+\texttt{trial\_end\_date} & 67.6\% $\pm$ 1.2\% & 50.0\% $\pm$ 6.5\% & 66.9\% $\pm$ 1.2\% & 44.1\% $\pm$ 6.5\% & 65.0\% $\pm$ 1.3\% & 12.3\% $\pm$ 4.3\% & 77.8\% $\pm$ 1.2\% & 13.5\% $\pm$ 4.7\% & 1487/59 (G: 1487/60) (H: 1321/57) (L: 1205/52) \\
+\texttt{trial\_start\_date} & 26.2\% $\pm$ 1.5\% & 9.7\% $\pm$ 1.2\% & 41.6\% $\pm$ 1.6\% & 61.3\% $\pm$ 1.9\% & 13.3\% $\pm$ 1.2\% & 4.4\% $\pm$ 0.8\% & 27.2\% $\pm$ 1.6\% & 3.6\% $\pm$ 0.8\% & 916/630 (G: 916/631) (H: 788/590) (L: 734/523) \\
+\bottomrule
+\end{tabular}%
+}
+\end{table*}

data/analysis/per_column.csv

@@ -0,0 +1,67 @@
+model,column,tp,mismatch,missed,hallucinated,tn,accuracy,recall_when_filled,precision_when_emitted,hallucination_rate,miss_rate,wrong_when_both_filled,f1
+gemini/gemini-3.1-flash-lite,court_cost_awarded_nominal,320,23,158,43,1003,0.8552,0.6387,0.8290,0.0411,0.3154,0.0671,0.7215
+gemini/gemini-3.1-flash-lite,defendant_no1_ISIC1_industry_category,623,298,154,311,161,0.5068,0.5795,0.5057,0.6589,0.1433,0.3236,0.5401
+gemini/gemini-3.1-flash-lite,defendants_all_count,1158,315,32,38,4,0.7511,0.7694,0.7664,0.9048,0.0213,0.2138,0.7679
+gemini/gemini-3.1-flash-lite,dispute_value_nominal,399,107,202,369,470,0.5617,0.5636,0.4560,0.4398,0.2853,0.2115,0.5041
+gemini/gemini-3.1-flash-lite,legal_subject_judgement,55,1482,4,5,1,0.0362,0.0357,0.0357,0.8333,0.0026,0.9642,0.0357
+gemini/gemini-3.1-flash-lite,party_compensation_awarded_nominal,260,13,91,78,1105,0.8824,0.7143,0.7407,0.0659,0.2500,0.0476,0.7273
+gemini/gemini-3.1-flash-lite,plaintiff_loosing_share,943,247,133,82,142,0.7014,0.7128,0.7414,0.3661,0.1005,0.2076,0.7268
+gemini/gemini-3.1-flash-lite,plaintiff_no1_ISIC1_industry_category,468,309,88,542,140,0.3930,0.5410,0.3548,0.7947,0.1017,0.3977,0.4286
+gemini/gemini-3.1-flash-lite,plaintiffs_all_count,1360,151,21,15,0,0.8791,0.8877,0.8912,1.0000,0.0137,0.0999,0.8895
+gemini/gemini-3.1-flash-lite,trial_end_date,1005,436,46,30,30,0.6690,0.6759,0.6832,0.5000,0.0309,0.3026,0.6795
+gemini/gemini-3.1-flash-lite,trial_start_date,240,107,569,61,570,0.5236,0.2620,0.5882,0.0967,0.6212,0.3084,0.3625
+gpt-5.4-mini,court_cost_awarded_nominal,328,35,138,87,958,0.8318,0.6547,0.7289,0.0833,0.2754,0.0964,0.6898
+gpt-5.4-mini,defendant_no1_ISIC1_industry_category,696,287,92,352,119,0.5272,0.6474,0.5213,0.7473,0.0856,0.2920,0.5776
+gpt-5.4-mini,defendants_all_count,1111,249,145,23,18,0.7303,0.7382,0.8033,0.5610,0.0963,0.1831,0.7694
+gpt-5.4-mini,dispute_value_nominal,404,100,204,451,387,0.5116,0.5706,0.4230,0.5382,0.2881,0.1984,0.4859
+gpt-5.4-mini,legal_subject_judgement,75,1458,7,6,0,0.0485,0.0487,0.0487,1.0000,0.0045,0.9511,0.0487
+gpt-5.4-mini,party_compensation_awarded_nominal,245,32,87,161,1021,0.8189,0.6731,0.5594,0.1362,0.2390,0.1155,0.6110
+gpt-5.4-mini,plaintiff_loosing_share,925,301,97,123,100,0.6630,0.6992,0.6857,0.5516,0.0733,0.2455,0.6924
+gpt-5.4-mini,plaintiff_no1_ISIC1_industry_category,466,308,91,548,133,0.3875,0.5387,0.3525,0.8047,0.1052,0.3979,0.4262
+gpt-5.4-mini,plaintiffs_all_count,1308,139,85,10,4,0.8486,0.8538,0.8977,0.7143,0.0555,0.0961,0.8752
+gpt-5.4-mini,trial_end_date,995,378,114,26,33,0.6649,0.6691,0.7112,0.4407,0.0767,0.2753,0.6895
+gpt-5.4-mini,trial_start_date,381,410,125,386,244,0.4043,0.4159,0.3237,0.6127,0.1365,0.5183,0.3641
+harvey,court_cost_awarded_nominal,247,25,155,18,933,0.8563,0.5785,0.8517,0.0189,0.3630,0.0919,0.6890
+harvey,defendant_no1_ISIC1_industry_category,462,202,274,164,276,0.5356,0.4925,0.5580,0.3727,0.2921,0.3042,0.5232
+harvey,defendants_all_count,961,153,229,17,18,0.7104,0.7156,0.8497,0.4857,0.1705,0.1373,0.7769
+harvey,dispute_value_nominal,278,74,239,259,528,0.5849,0.4704,0.4550,0.3291,0.4044,0.2102,0.4626
+harvey,legal_subject_judgement,53,1052,267,5,1,0.0392,0.0386,0.0477,0.8333,0.1946,0.9520,0.0427
+harvey,party_compensation_awarded_nominal,190,5,89,51,1043,0.8948,0.6690,0.7724,0.0466,0.3134,0.0256,0.7170
+harvey,plaintiff_loosing_share,849,195,145,63,126,0.7075,0.7140,0.7669,0.3333,0.1220,0.1868,0.7395
+harvey,plaintiff_no1_ISIC1_industry_category,314,165,273,159,467,0.5668,0.4176,0.4922,0.2540,0.3630,0.3445,0.4518
+harvey,plaintiffs_all_count,1023,121,222,10,2,0.7438,0.7489,0.8865,0.8333,0.1625,0.1058,0.8119
+harvey,trial_end_date,859,70,392,7,50,0.6597,0.6503,0.9177,0.1228,0.2967,0.0753,0.7612
+harvey,trial_start_date,105,53,630,26,564,0.4855,0.1332,0.5707,0.0441,0.7995,0.3354,0.2160
+harvey-2,court_cost_awarded_nominal,247,40,146,8,952,0.8607,0.5704,0.8373,0.0083,0.3372,0.1394,0.6786
+harvey-2,defendant_no1_ISIC1_industry_category,320,185,446,121,321,0.4602,0.3365,0.5112,0.2738,0.4690,0.3663,0.4058
+harvey-2,defendants_all_count,974,154,228,23,14,0.7093,0.7183,0.8462,0.6216,0.1681,0.1365,0.7770
+harvey-2,dispute_value_nominal,201,69,338,160,625,0.5930,0.3306,0.4674,0.2038,0.5559,0.2556,0.3873
+harvey-2,legal_subject_judgement,10,778,599,0,6,0.0115,0.0072,0.0127,0.0000,0.4319,0.9873,0.0092
+harvey-2,party_compensation_awarded_nominal,183,4,103,54,1049,0.8844,0.6310,0.7593,0.0490,0.3552,0.0214,0.6893
+harvey-2,plaintiff_loosing_share,831,112,249,32,169,0.7179,0.6971,0.8523,0.1592,0.2089,0.1188,0.7670
+harvey-2,plaintiff_no1_ISIC1_industry_category,236,134,384,120,519,0.5420,0.3130,0.4816,0.1878,0.5093,0.3622,0.3794
+harvey-2,plaintiffs_all_count,1072,116,193,8,4,0.7724,0.7762,0.8963,0.6667,0.1398,0.0976,0.8320
+harvey-2,trial_end_date,868,134,334,8,49,0.6583,0.6497,0.8594,0.1404,0.2500,0.1337,0.7400
+harvey-2,trial_start_date,118,37,627,20,591,0.5090,0.1509,0.6743,0.0327,0.8018,0.2387,0.2466
+legora-1,court_cost_awarded_nominal,216,3,89,9,940,0.9196,0.7013,0.9474,0.0095,0.2890,0.0137,0.8060
+legora-1,defendant_no1_ISIC1_industry_category,485,116,273,86,297,0.6221,0.5549,0.7060,0.2245,0.3124,0.1930,0.6214
+legora-1,defendants_all_count,814,123,285,13,22,0.6651,0.6661,0.8568,0.3714,0.2332,0.1313,0.7495
+legora-1,dispute_value_nominal,187,22,330,88,630,0.6500,0.3469,0.6296,0.1226,0.6122,0.1053,0.4474
+legora-1,legal_subject_judgement,23,735,493,1,5,0.0223,0.0184,0.0303,0.1667,0.3941,0.9697,0.0229
+legora-1,party_compensation_awarded_nominal,186,3,46,43,979,0.9268,0.7915,0.8017,0.0421,0.1957,0.0159,0.7966
+legora-1,plaintiff_loosing_share,755,61,274,19,148,0.7184,0.6927,0.9042,0.1138,0.2514,0.0748,0.7844
+legora-1,plaintiff_no1_ISIC1_industry_category,251,117,314,97,478,0.5800,0.3680,0.5398,0.1687,0.4604,0.3179,0.4377
+legora-1,plaintiffs_all_count,876,47,320,4,10,0.7049,0.7047,0.9450,0.2857,0.2574,0.0509,0.8074
+legora-1,trial_end_date,937,123,145,7,45,0.7812,0.7776,0.8782,0.1346,0.1203,0.1160,0.8248
+legora-1,trial_start_date,200,35,499,19,504,0.5601,0.2725,0.7874,0.0363,0.6798,0.1489,0.4049
+legora-2,court_cost_awarded_nominal,221,5,88,7,950,0.9213,0.7038,0.9485,0.0073,0.2803,0.0221,0.8080
+legora-2,defendant_no1_ISIC1_industry_category,447,129,305,111,279,0.5712,0.5074,0.6507,0.2846,0.3462,0.2240,0.5702
+legora-2,defendants_all_count,811,98,324,7,31,0.6625,0.6577,0.8854,0.1842,0.2628,0.1078,0.7548
+legora-2,dispute_value_nominal,218,34,297,93,629,0.6664,0.3971,0.6319,0.1288,0.5410,0.1349,0.4877
+legora-2,legal_subject_judgement,28,719,518,1,5,0.0260,0.0221,0.0374,0.1667,0.4095,0.9625,0.0278
+legora-2,party_compensation_awarded_nominal,192,3,46,46,984,0.9253,0.7967,0.7967,0.0447,0.1909,0.0154,0.7967
+legora-2,plaintiff_loosing_share,778,103,218,28,144,0.7254,0.7079,0.8559,0.1628,0.1984,0.1169,0.7749
+legora-2,plaintiff_no1_ISIC1_industry_category,252,113,318,87,501,0.5924,0.3690,0.5575,0.1480,0.4656,0.3096,0.4441
+legora-2,plaintiffs_all_count,910,59,288,5,9,0.7231,0.7239,0.9343,0.3571,0.2291,0.0609,0.8158
+legora-2,trial_end_date,938,118,162,7,46,0.7742,0.7701,0.8824,0.1321,0.1330,0.1117,0.8224
+legora-2,trial_start_date,201,29,510,19,512,0.5610,0.2716,0.8072,0.0358,0.6892,0.1261,0.4065

data/analysis/per_country.csv

@@ -0,0 +1,115 @@
+country,model,legal_tradition,language_family,tp,mismatch,missed,hallucinated,tn,accuracy,recall_when_filled,precision_when_emitted,hallucination_rate,miss_rate,wrong_when_both_filled,f1,cost_tp,cost_mismatch,cost_missed,cost_hallucinated,cost_tn,cost_accuracy,cost_recall_when_filled,cost_precision_when_emitted,cost_hallucination_rate,cost_miss_rate,cost_wrong_when_both_filled,cost_f1
+am,gemini/gemini-3.1-flash-lite,civil,non-latin,259,77,179,13,110,0.5784,0.5029,0.7421,0.1057,0.3476,0.2292,0.5995,41,2,104,0,85,0.5431,0.2789,0.9535,0.0000,0.7075,0.0465,0.4316
+am,gpt-5.4-mini,civil,non-latin,277,83,155,19,104,0.5972,0.5379,0.7309,0.1545,0.3010,0.2306,0.6197,46,3,98,0,85,0.5647,0.3129,0.9388,0.0000,0.6667,0.0612,0.4694
+am,harvey,civil,non-latin,42,66,397,0,122,0.2616,0.0832,0.3889,0.0000,0.7861,0.6111,0.1370,38,0,105,0,85,0.5395,0.2657,1.0000,0.0000,0.7343,0.0000,0.4199
+am,harvey-2,civil,non-latin,105,84,316,2,120,0.3589,0.2079,0.5497,0.0164,0.6257,0.4444,0.3017,38,0,105,0,85,0.5395,0.2657,1.0000,0.0000,0.7343,0.0000,0.4199
+am,legora-1,civil,non-latin,0,0,0,0,0,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0,0,0,0,0,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+am,legora-2,civil,non-latin,0,0,0,0,0,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0,0,0,0,0,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+au,gemini/gemini-3.1-flash-lite,common,en-latin,132,54,7,47,90,0.6727,0.6839,0.5665,0.3431,0.0363,0.2903,0.6197,22,6,3,16,73,0.7917,0.7097,0.5000,0.1798,0.0968,0.2143,0.5867
+au,gpt-5.4-mini,common,en-latin,130,59,4,69,68,0.6000,0.6736,0.5039,0.5036,0.0207,0.3122,0.5765,22,7,2,27,62,0.7000,0.7097,0.3929,0.3034,0.0645,0.2414,0.5057
+au,harvey,common,en-latin,135,44,14,29,108,0.7364,0.6995,0.6490,0.2117,0.0725,0.2458,0.6733,19,10,2,17,72,0.7583,0.6129,0.4130,0.1910,0.0645,0.3448,0.4935
+au,harvey-2,common,en-latin,129,47,17,15,122,0.7606,0.6684,0.6754,0.1095,0.0881,0.2670,0.6719,21,5,5,8,81,0.8500,0.6774,0.6176,0.0899,0.1613,0.1923,0.6462
+au,legora-1,common,en-latin,115,43,35,7,130,0.7424,0.5959,0.6970,0.0511,0.1813,0.2722,0.6425,19,7,5,2,87,0.8833,0.6129,0.6786,0.0225,0.1613,0.2692,0.6441
+au,legora-2,common,en-latin,118,39,36,17,120,0.7212,0.6114,0.6782,0.1241,0.1865,0.2484,0.6431,18,8,5,6,83,0.8417,0.5806,0.5625,0.0674,0.1613,0.3077,0.5714
+be,gemini/gemini-3.1-flash-lite,civil,eu-latin,278,82,50,19,176,0.7504,0.6780,0.7335,0.0974,0.1220,0.2278,0.7047,47,8,23,12,130,0.8045,0.6026,0.7015,0.0845,0.2949,0.1455,0.6483
+be,gpt-5.4-mini,civil,eu-latin,261,118,31,53,142,0.6661,0.6366,0.6042,0.2718,0.0756,0.3113,0.6200,39,23,16,18,124,0.7409,0.5000,0.4875,0.1268,0.2051,0.3710,0.4937
+be,harvey,civil,eu-latin,254,89,67,16,179,0.7157,0.6195,0.7075,0.0821,0.1634,0.2595,0.6606,47,14,17,14,128,0.7955,0.6026,0.6267,0.0986,0.2179,0.2295,0.6144
+be,harvey-2,civil,eu-latin,224,68,118,6,189,0.6826,0.5463,0.7517,0.0308,0.2878,0.2329,0.6328,39,6,33,5,137,0.8000,0.5000,0.7800,0.0352,0.4231,0.1333,0.6094
+be,legora-1,civil,eu-latin,235,62,113,5,190,0.7025,0.5732,0.7781,0.0256,0.2756,0.2088,0.6601,43,2,33,4,138,0.8227,0.5513,0.8776,0.0282,0.4231,0.0444,0.6772
+be,legora-2,civil,eu-latin,238,65,107,4,191,0.7091,0.5805,0.7752,0.0205,0.2610,0.2145,0.6639,46,4,28,3,139,0.8409,0.5897,0.8679,0.0211,0.3590,0.0800,0.7023
+br,gemini/gemini-3.1-flash-lite,civil,eu-latin,365,230,98,206,531,0.6266,0.5267,0.4557,0.2795,0.1414,0.3866,0.4886,96,19,22,32,351,0.8596,0.7007,0.6531,0.0836,0.1606,0.1652,0.6761
+br,gpt-5.4-mini,civil,eu-latin,321,243,129,265,472,0.5545,0.4632,0.3872,0.3596,0.1861,0.4309,0.4218,99,26,12,79,304,0.7750,0.7226,0.4853,0.2063,0.0876,0.2080,0.5806
+br,harvey,civil,eu-latin,280,213,200,162,575,0.5979,0.4040,0.4275,0.2198,0.2886,0.4320,0.4154,73,21,43,36,347,0.8077,0.5328,0.5615,0.0940,0.3139,0.2234,0.5468
+br,harvey-2,civil,eu-latin,266,169,248,107,618,0.6278,0.3895,0.4908,0.1476,0.3631,0.3885,0.4343,68,12,55,15,362,0.8398,0.5037,0.7158,0.0398,0.4074,0.1500,0.5913
+br,legora-1,civil,eu-latin,262,150,250,82,620,0.6466,0.3958,0.5304,0.1168,0.3776,0.3641,0.4533,65,9,57,6,359,0.8548,0.4962,0.8125,0.0164,0.4351,0.1216,0.6161
+br,legora-2,civil,eu-latin,269,152,249,94,611,0.6400,0.4015,0.5223,0.1333,0.3716,0.3610,0.4540,70,13,51,10,356,0.8520,0.5224,0.7527,0.0273,0.3806,0.1566,0.6167
+ch,gemini/gemini-3.1-flash-lite,civil,eu-latin,887,184,48,133,178,0.7448,0.7927,0.7367,0.4277,0.0429,0.1718,0.7637,370,21,7,28,94,0.8923,0.9296,0.8831,0.2295,0.0176,0.0537,0.9058
+ch,gpt-5.4-mini,civil,eu-latin,919,192,8,247,64,0.6874,0.8213,0.6767,0.7942,0.0071,0.1728,0.7420,373,21,4,81,41,0.7962,0.9372,0.7853,0.6639,0.0101,0.0533,0.8545
+ch,harvey,civil,eu-latin,878,190,51,50,261,0.7965,0.7846,0.7853,0.1608,0.0456,0.1779,0.7850,343,35,20,20,102,0.8558,0.8618,0.8618,0.1639,0.0503,0.0926,0.8618
+ch,harvey-2,civil,eu-latin,849,179,91,33,278,0.7881,0.7587,0.8002,0.1061,0.0813,0.1741,0.7789,324,20,54,16,106,0.8269,0.8141,0.9000,0.1311,0.1357,0.0581,0.8549
+ch,legora-1,civil,eu-latin,895,169,55,24,287,0.8266,0.7998,0.8226,0.0772,0.0492,0.1588,0.8111,354,16,28,16,106,0.8846,0.8894,0.9171,0.1311,0.0704,0.0432,0.9031
+ch,legora-2,civil,eu-latin,901,175,43,24,287,0.8308,0.8052,0.8191,0.0772,0.0384,0.1626,0.8121,353,27,18,15,107,0.8846,0.8869,0.8937,0.1230,0.0452,0.0711,0.8903
+de,gemini/gemini-3.1-flash-lite,civil,eu-latin,367,367,38,191,467,0.5832,0.4754,0.3968,0.2903,0.0492,0.5000,0.4325,134,30,17,45,294,0.8231,0.7403,0.6411,0.1327,0.0939,0.1829,0.6872
+de,gpt-5.4-mini,civil,eu-latin,352,333,87,217,441,0.5545,0.4560,0.3902,0.3298,0.1127,0.4861,0.4205,137,25,19,46,293,0.8269,0.7569,0.6587,0.1357,0.1050,0.1543,0.7044
+de,harvey,civil,eu-latin,339,144,275,103,536,0.6263,0.4472,0.5785,0.1612,0.3628,0.2981,0.5045,131,29,18,41,289,0.8268,0.7360,0.6517,0.1242,0.1011,0.1812,0.6913
+de,harvey-2,civil,eu-latin,318,139,305,57,589,0.6442,0.4173,0.6187,0.0882,0.4003,0.3042,0.4984,109,25,46,22,310,0.8184,0.6056,0.6987,0.0663,0.2556,0.1866,0.6488
+de,legora-1,civil,eu-latin,138,142,437,49,532,0.5162,0.1925,0.4195,0.0843,0.6095,0.5071,0.2639,112,20,38,19,283,0.8369,0.6588,0.7417,0.0629,0.2235,0.1515,0.6978
+de,legora-2,civil,eu-latin,148,144,441,51,558,0.5261,0.2019,0.4315,0.0837,0.6016,0.4932,0.2751,117,20,37,21,293,0.8402,0.6724,0.7405,0.0669,0.2126,0.1460,0.7048
+es,gemini/gemini-3.1-flash-lite,civil,eu-latin,809,300,89,68,164,0.6804,0.6753,0.6873,0.2931,0.0743,0.2705,0.6813,244,47,32,36,161,0.7788,0.7554,0.7462,0.1827,0.0991,0.1615,0.7508
+es,gpt-5.4-mini,civil,eu-latin,831,277,90,103,129,0.6713,0.6937,0.6862,0.4440,0.0751,0.2500,0.6899,248,46,29,72,125,0.7173,0.7678,0.6776,0.3655,0.0898,0.1565,0.7199
+es,harvey,civil,eu-latin,68,29,45,6,17,0.5152,0.4789,0.6602,0.2609,0.3169,0.2990,0.5551,31,4,10,2,13,0.7333,0.6889,0.8378,0.1333,0.2222,0.1143,0.7561
+es,harvey-2,civil,eu-latin,72,18,52,4,19,0.5515,0.5070,0.7660,0.1739,0.3662,0.2000,0.6102,30,4,11,1,14,0.7333,0.6667,0.8571,0.0667,0.2444,0.1176,0.7500
+es,legora-1,civil,eu-latin,626,117,455,25,207,0.5825,0.5225,0.8151,0.1078,0.3798,0.1575,0.6368,205,13,105,7,190,0.7596,0.6347,0.9111,0.0355,0.3251,0.0596,0.7482
+es,legora-2,civil,eu-latin,627,132,439,24,208,0.5839,0.5234,0.8008,0.1034,0.3664,0.1739,0.6330,202,23,98,6,191,0.7558,0.6254,0.8745,0.0305,0.3034,0.1022,0.7292
+fr,gemini/gemini-3.1-flash-lite,civil,eu-latin,135,58,26,43,68,0.6152,0.6164,0.5720,0.3874,0.1187,0.3005,0.5934,29,8,16,30,37,0.5500,0.5472,0.4328,0.4478,0.3019,0.2162,0.4833
+fr,gpt-5.4-mini,civil,eu-latin,154,48,17,70,41,0.5909,0.7032,0.5662,0.6306,0.0776,0.2376,0.6273,31,6,16,28,39,0.5833,0.5849,0.4769,0.4179,0.3019,0.1622,0.5254
+fr,harvey,civil,eu-latin,131,45,43,35,76,0.6273,0.5982,0.6209,0.3153,0.1963,0.2557,0.6093,30,3,20,26,41,0.5917,0.5660,0.5085,0.3881,0.3774,0.0909,0.5357
+fr,harvey-2,civil,eu-latin,125,30,64,29,82,0.6273,0.5708,0.6793,0.2613,0.2922,0.1935,0.6203,35,2,16,28,39,0.6167,0.6604,0.5385,0.4179,0.3019,0.0541,0.5932
+fr,legora-1,civil,eu-latin,126,29,64,28,83,0.6333,0.5753,0.6885,0.2523,0.2922,0.1871,0.6269,33,1,19,25,42,0.6250,0.6226,0.5593,0.3731,0.3585,0.0294,0.5893
+fr,legora-2,civil,eu-latin,126,27,66,29,82,0.6303,0.5753,0.6923,0.2613,0.3014,0.1765,0.6284,35,1,17,28,39,0.6167,0.6604,0.5469,0.4179,0.3208,0.0278,0.5983
+ge,gemini/gemini-3.1-flash-lite,civil,non-latin,494,289,225,58,166,0.5357,0.4901,0.5874,0.2589,0.2232,0.3691,0.5343,110,55,111,48,124,0.5223,0.3986,0.5164,0.2791,0.4022,0.3333,0.4499
+ge,gpt-5.4-mini,civil,non-latin,495,356,157,111,113,0.4935,0.4911,0.5146,0.4955,0.1558,0.4183,0.5025,92,88,96,73,99,0.4263,0.3333,0.3636,0.4244,0.3478,0.4889,0.3478
+ge,harvey,civil,non-latin,359,206,443,36,188,0.4440,0.3562,0.5973,0.1607,0.4395,0.3646,0.4462,109,37,130,34,138,0.5513,0.3949,0.6056,0.1977,0.4710,0.2534,0.4781
+ge,harvey-2,civil,non-latin,358,217,433,24,200,0.4529,0.3552,0.5977,0.1071,0.4296,0.3774,0.4456,105,65,106,22,150,0.5692,0.3804,0.5469,0.1279,0.3841,0.3824,0.4487
+ge,legora-1,civil,non-latin,0,0,0,0,0,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0,0,0,0,0,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+ge,legora-2,civil,non-latin,0,0,0,0,0,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0,0,0,0,0,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+hk,gemini/gemini-3.1-flash-lite,common,en-latin,46,23,3,12,26,0.6545,0.6389,0.5679,0.3158,0.0417,0.3333,0.6013,5,6,1,5,23,0.7000,0.4167,0.3125,0.1786,0.0833,0.5455,0.3571
+hk,gpt-5.4-mini,common,en-latin,44,21,7,11,27,0.6455,0.6111,0.5789,0.2895,0.0972,0.3231,0.5946,3,5,4,2,26,0.7250,0.2500,0.3000,0.0714,0.3333,0.6250,0.2727
+hk,harvey,common,en-latin,38,17,17,4,34,0.6545,0.5278,0.6441,0.1053,0.2361,0.3091,0.5802,5,4,3,1,27,0.8000,0.4167,0.5000,0.0357,0.2500,0.4444,0.4545
+hk,harvey-2,common,en-latin,28,6,26,5,23,0.5795,0.4667,0.7179,0.1786,0.4333,0.1765,0.5657,3,2,5,4,18,0.6562,0.3000,0.3333,0.1818,0.5000,0.4000,0.3158
+hk,legora-1,common,en-latin,39,6,27,1,37,0.6909,0.5417,0.8478,0.0263,0.3750,0.1333,0.6610,4,2,6,0,28,0.8000,0.3333,0.6667,0.0000,0.5000,0.3333,0.4444
+hk,legora-2,common,en-latin,35,8,29,3,35,0.6364,0.4861,0.7609,0.0789,0.4028,0.1860,0.5932,2,3,7,1,27,0.7250,0.1667,0.3333,0.0357,0.5833,0.6000,0.2222
+in,gemini/gemini-3.1-flash-lite,common,en-latin,100,51,6,36,60,0.6324,0.6369,0.5348,0.3750,0.0382,0.3377,0.5814,32,9,4,14,33,0.7065,0.7111,0.5818,0.2979,0.0889,0.2195,0.6400
+in,gpt-5.4-mini,common,en-latin,110,41,6,57,39,0.5889,0.7006,0.5288,0.5938,0.0382,0.2715,0.6027,35,7,3,14,33,0.7391,0.7778,0.6250,0.2979,0.0667,0.1667,0.6931
+in,harvey,common,en-latin,113,34,18,20,79,0.7273,0.6848,0.6766,0.2020,0.1091,0.2313,0.6807,28,9,9,8,42,0.7292,0.6087,0.6222,0.1600,0.1957,0.2432,0.6154
+in,harvey-2,common,en-latin,109,37,19,9,90,0.7538,0.6606,0.7032,0.0909,0.1152,0.2534,0.6813,36,4,6,2,48,0.8750,0.7826,0.8571,0.0400,0.1304,0.1000,0.8182
+in,legora-1,common,en-latin,111,23,31,10,89,0.7576,0.6727,0.7708,0.1010,0.1879,0.1716,0.7184,31,0,15,0,50,0.8438,0.6739,1.0000,0.0000,0.3261,0.0000,0.8052
+in,legora-2,common,en-latin,108,29,28,11,88,0.7424,0.6545,0.7297,0.1111,0.1697,0.2117,0.6901,30,2,14,4,46,0.7917,0.6522,0.8333,0.0800,0.3043,0.0625,0.7317
+np,gemini/gemini-3.1-flash-lite,civil,non-latin,541,426,95,236,132,0.4706,0.5094,0.4497,0.6413,0.0895,0.4405,0.4777,272,68,25,81,74,0.6654,0.7452,0.6461,0.5226,0.0685,0.2000,0.6921
+np,gpt-5.4-mini,civil,non-latin,568,457,37,287,81,0.4538,0.5348,0.4329,0.7799,0.0348,0.4459,0.4785,274,72,19,103,52,0.6269,0.7507,0.6102,0.6645,0.0521,0.2081,0.6732
+np,harvey,civil,non-latin,240,132,326,54,150,0.4324,0.3438,0.5634,0.2647,0.4670,0.3548,0.4270,200,8,38,11,71,0.8262,0.8130,0.9132,0.1341,0.1545,0.0385,0.8602
+np,harvey-2,civil,non-latin,224,101,490,70,215,0.3991,0.2748,0.5671,0.2456,0.6012,0.3108,0.3702,216,3,61,1,119,0.8375,0.7714,0.9818,0.0083,0.2179,0.0137,0.8640
+np,legora-1,civil,non-latin,69,19,142,14,64,0.4318,0.3000,0.6765,0.1795,0.6174,0.2159,0.4157,69,1,10,8,24,0.8304,0.8625,0.8846,0.2500,0.1250,0.0143,0.8734
+np,legora-2,civil,non-latin,91,20,192,18,86,0.4349,0.3003,0.7054,0.1731,0.6337,0.1802,0.4213,85,2,17,6,38,0.8311,0.8173,0.9140,0.1364,0.1635,0.0230,0.8629
+nz,gemini/gemini-3.1-flash-lite,common,en-latin,144,41,2,43,89,0.7304,0.7701,0.6316,0.3258,0.0107,0.2216,0.6940,35,2,0,14,65,0.8621,0.9459,0.6863,0.1772,0.0000,0.0541,0.7955
+nz,gpt-5.4-mini,common,en-latin,136,49,2,68,64,0.6270,0.7273,0.5375,0.5152,0.0107,0.2649,0.6182,27,8,2,25,54,0.6983,0.7297,0.4500,0.3165,0.0541,0.2286,0.5567
+nz,harvey,common,en-latin,141,11,35,19,113,0.7962,0.7540,0.8246,0.1439,0.1872,0.0724,0.7877,33,3,1,17,62,0.8190,0.8919,0.6226,0.2152,0.0270,0.0833,0.7333
+nz,harvey-2,common,en-latin,138,6,43,15,117,0.7994,0.7380,0.8679,0.1136,0.2299,0.0417,0.7977,30,3,4,13,66,0.8276,0.8108,0.6522,0.1646,0.1081,0.0909,0.7229
+nz,legora-1,common,en-latin,136,7,44,8,124,0.8150,0.7273,0.9007,0.0606,0.2353,0.0490,0.8047,29,1,7,6,73,0.8793,0.7838,0.8056,0.0759,0.1892,0.0333,0.7945
+nz,legora-2,common,en-latin,143,6,38,9,123,0.8339,0.7647,0.9051,0.0682,0.2032,0.0403,0.8290,35,0,2,5,74,0.9397,0.9459,0.8750,0.0633,0.0541,0.0000,0.9091
+ph,gemini/gemini-3.1-flash-lite,common,en-latin,24,38,3,19,26,0.4545,0.3692,0.2963,0.4222,0.0462,0.6129,0.3288,8,7,1,7,17,0.6250,0.5000,0.3636,0.2917,0.0625,0.4667,0.4211
+ph,gpt-5.4-mini,common,en-latin,32,29,4,22,23,0.5000,0.4923,0.3855,0.4889,0.0615,0.4754,0.4324,13,3,0,7,17,0.7500,0.8125,0.5652,0.2917,0.0000,0.1875,0.6667
+ph,harvey,common,en-latin,39,14,10,10,26,0.6566,0.6190,0.6190,0.2778,0.1587,0.2642,0.6190,10,5,1,6,14,0.6667,0.6250,0.4762,0.3000,0.0625,0.3333,0.5405
+ph,harvey-2,common,en-latin,42,7,14,10,26,0.6869,0.6667,0.7119,0.2778,0.2222,0.1429,0.6885,12,2,2,5,15,0.7500,0.7500,0.6316,0.2500,0.1250,0.1429,0.6857
+ph,legora-1,common,en-latin,43,7,15,11,34,0.7000,0.6615,0.7049,0.2444,0.2308,0.1400,0.6825,9,1,6,4,20,0.7250,0.5625,0.6429,0.1667,0.3750,0.1000,0.6000
+ph,legora-2,common,en-latin,45,10,10,10,35,0.7273,0.6923,0.6923,0.2222,0.1538,0.1818,0.6923,11,1,4,4,20,0.7750,0.6875,0.6875,0.1667,0.2500,0.0833,0.6875
+rs,gemini/gemini-3.1-flash-lite,civil,eu-latin,138,43,5,40,60,0.6923,0.7419,0.6244,0.4000,0.0269,0.2376,0.6781,38,6,4,18,38,0.7308,0.7917,0.6129,0.3214,0.0833,0.1364,0.6909
+rs,gpt-5.4-mini,civil,eu-latin,137,41,8,36,64,0.7028,0.7366,0.6402,0.3600,0.0430,0.2303,0.6850,38,3,7,7,49,0.8365,0.7917,0.7917,0.1250,0.1458,0.0732,0.7917
+rs,harvey,civil,eu-latin,131,37,11,10,86,0.7891,0.7318,0.7360,0.1042,0.0615,0.2202,0.7339,30,10,6,7,47,0.7700,0.6522,0.6383,0.1296,0.1304,0.2500,0.6452
+rs,harvey-2,civil,eu-latin,113,20,46,12,84,0.7164,0.6313,0.7793,0.1250,0.2570,0.1504,0.6975,27,7,12,12,42,0.6900,0.5870,0.5870,0.2222,0.2609,0.2059,0.5870
+rs,legora-1,civil,eu-latin,127,11,48,12,88,0.7517,0.6828,0.8467,0.1200,0.2581,0.0797,0.7560,26,3,19,12,44,0.6731,0.5417,0.6341,0.2143,0.3958,0.1034,0.5843
+rs,legora-2,civil,eu-latin,128,22,36,9,91,0.7657,0.6882,0.8050,0.0900,0.1935,0.1467,0.7420,27,6,15,8,48,0.7212,0.5625,0.6585,0.1429,0.3125,0.1818,0.6067
+sg,gemini/gemini-3.1-flash-lite,common,en-latin,503,243,233,93,292,0.5828,0.5138,0.5995,0.2416,0.2380,0.3257,0.5534,86,21,78,45,266,0.7097,0.4649,0.5658,0.1447,0.4216,0.1963,0.5104
+sg,gpt-5.4-mini,common,en-latin,505,346,127,113,262,0.5669,0.5164,0.5239,0.3013,0.1299,0.4066,0.5201,78,35,72,55,252,0.6707,0.4216,0.4643,0.1792,0.3892,0.3097,0.4419
+sg,harvey,common,en-latin,495,199,284,45,330,0.6098,0.5061,0.6698,0.1200,0.2904,0.2867,0.5766,77,19,89,22,285,0.7358,0.4162,0.6525,0.0717,0.4811,0.1979,0.5083
+sg,harvey-2,common,en-latin,453,183,342,42,333,0.5809,0.4632,0.6681,0.1120,0.3497,0.2877,0.5471,61,11,113,21,286,0.7053,0.3297,0.6559,0.0684,0.6108,0.1528,0.4388
+sg,legora-1,common,en-latin,406,163,409,19,356,0.5632,0.4151,0.6905,0.0507,0.4182,0.2865,0.5185,40,0,145,2,305,0.7012,0.2162,0.9524,0.0065,0.7838,0.0000,0.3524
+sg,legora-2,common,en-latin,426,158,394,18,357,0.5787,0.4356,0.7076,0.0480,0.4029,0.2705,0.5392,50,4,131,3,304,0.7195,0.2703,0.8772,0.0098,0.7081,0.0741,0.4132
+tw,gemini/gemini-3.1-flash-lite,civil,non-latin,567,186,189,144,344,0.6371,0.6019,0.6321,0.2951,0.2006,0.2470,0.6166,143,7,68,41,261,0.7769,0.6560,0.7487,0.1358,0.3119,0.0467,0.6993
+tw,gpt-5.4-mini,civil,non-latin,601,226,115,213,275,0.6126,0.6380,0.5779,0.4365,0.1221,0.2733,0.6065,143,7,68,75,227,0.7115,0.6560,0.6356,0.2483,0.3119,0.0467,0.6456
+tw,harvey,civil,non-latin,520,181,241,48,440,0.6713,0.5520,0.6943,0.0984,0.2558,0.2582,0.6150,121,20,77,32,270,0.7519,0.5550,0.6994,0.1060,0.3532,0.1418,0.6189
+tw,harvey-2,civil,non-latin,484,154,304,29,459,0.6594,0.5138,0.7256,0.0594,0.3227,0.2414,0.6016,102,7,109,18,284,0.7423,0.4679,0.8031,0.0596,0.5000,0.0642,0.5913
+tw,legora-1,civil,non-latin,550,157,235,23,465,0.7098,0.5839,0.7534,0.0471,0.2495,0.2221,0.6579,141,2,75,17,285,0.8192,0.6468,0.8812,0.0563,0.3440,0.0140,0.7460
+tw,legora-2,civil,non-latin,549,159,234,28,460,0.7056,0.5828,0.7459,0.0574,0.2484,0.2246,0.6544,144,3,71,17,285,0.8250,0.6606,0.8780,0.0563,0.3257,0.0204,0.7539
+uk,gemini/gemini-3.1-flash-lite,common,en-latin,421,446,114,113,336,0.5294,0.4292,0.4296,0.2517,0.1162,0.5144,0.4294,88,25,26,83,298,0.7423,0.6331,0.4490,0.2178,0.1871,0.2212,0.5254
+uk,gpt-5.4-mini,common,en-latin,461,428,92,129,320,0.5462,0.4699,0.4528,0.2873,0.0938,0.4814,0.4612,82,30,27,78,303,0.7404,0.5899,0.4316,0.2047,0.1942,0.2679,0.4985
+uk,harvey,common,en-latin,575,258,148,99,350,0.6469,0.5861,0.6170,0.2205,0.1509,0.3097,0.6012,100,34,5,85,296,0.7615,0.7194,0.4566,0.2231,0.0360,0.2537,0.5587
+uk,harvey-2,common,en-latin,524,171,286,62,387,0.6371,0.5341,0.6922,0.1381,0.2915,0.2460,0.6030,101,20,18,54,327,0.8231,0.7266,0.5771,0.1417,0.1295,0.1653,0.6433
+uk,legora-1,common,en-latin,535,141,305,42,407,0.6587,0.5454,0.7451,0.0935,0.3109,0.2086,0.6298,76,3,60,29,352,0.8231,0.5468,0.7037,0.0761,0.4317,0.0380,0.6154
+uk,legora-2,common,en-latin,539,153,289,39,410,0.6636,0.5494,0.7373,0.0869,0.2946,0.2211,0.6297,88,14,37,32,349,0.8404,0.6331,0.6567,0.0840,0.2662,0.1373,0.6447
+us,gemini/gemini-3.1-flash-lite,common,en-latin,621,350,88,60,311,0.6517,0.5864,0.6023,0.1617,0.0831,0.3605,0.5943,122,43,42,17,296,0.8038,0.5894,0.6703,0.0543,0.2029,0.2606,0.6272
+us,gpt-5.4-mini,common,en-latin,600,350,109,83,288,0.6210,0.5666,0.5808,0.2237,0.1029,0.3684,0.5736,122,53,32,32,281,0.7750,0.5894,0.5894,0.1022,0.1546,0.3029,0.5894
+us,harvey,common,en-latin,563,206,290,33,338,0.6301,0.5316,0.7020,0.0889,0.2738,0.2679,0.6051,139,34,34,12,301,0.8462,0.6715,0.7514,0.0383,0.1643,0.1965,0.7092
+us,harvey-2,common,en-latin,499,127,433,23,348,0.5923,0.4712,0.7689,0.0620,0.4089,0.2029,0.5843,105,27,75,7,306,0.7904,0.5072,0.7554,0.0224,0.3623,0.2045,0.6069
+us,legora-1,common,en-latin,517,139,403,26,345,0.6028,0.4882,0.7581,0.0701,0.3805,0.2119,0.5939,88,8,111,2,311,0.7673,0.4251,0.8980,0.0064,0.5362,0.0833,0.5770
+us,legora-2,common,en-latin,505,111,443,23,348,0.5965,0.4769,0.7903,0.0620,0.4183,0.1802,0.5948,96,14,97,5,308,0.7769,0.4638,0.8348,0.0160,0.4686,0.1273,0.5963

data/analysis/per_language.csv

@@ -0,0 +1,19 @@
+model,group,n_countries,tp,mismatch,missed,hallucinated,tn,accuracy,recall_when_filled,precision_when_emitted,hallucination_rate,miss_rate,wrong_when_both_filled,f1,cost_tp,cost_mismatch,cost_missed,cost_hallucinated,cost_tn,cost_accuracy,cost_recall_when_filled,cost_precision_when_emitted,cost_hallucination_rate,cost_miss_rate,cost_wrong_when_both_filled,cost_f1
+gemini/gemini-3.1-flash-lite,en-latin,8,1991,1246,456,423,1230,0.6025,0.5391,0.5440,0.2559,0.1235,0.3849,0.5415,398,119,155,201,1071,0.7557,0.5923,0.5543,0.1580,0.2307,0.2302,0.5727
+gemini/gemini-3.1-flash-lite,eu-latin,7,2979,1264,354,700,1644,0.6660,0.6480,0.6027,0.2986,0.0770,0.2979,0.6245,958,139,121,201,1105,0.8174,0.7865,0.7381,0.1539,0.0993,0.1267,0.7615
+gemini/gemini-3.1-flash-lite,non-latin,4,1861,978,688,451,752,0.5524,0.5276,0.5657,0.3749,0.1951,0.3445,0.5460,566,132,308,170,544,0.6453,0.5626,0.6521,0.2381,0.3062,0.1891,0.6041
+gpt-5.4-mini,en-latin,8,2018,1323,351,552,1091,0.5828,0.5466,0.5184,0.3360,0.0951,0.3960,0.5321,382,148,142,240,1028,0.7268,0.5685,0.4961,0.1893,0.2113,0.2792,0.5298
+gpt-5.4-mini,eu-latin,7,2975,1252,370,991,1353,0.6235,0.6472,0.5701,0.4228,0.0805,0.2962,0.6062,965,150,103,331,975,0.7686,0.7923,0.6674,0.2534,0.0846,0.1345,0.7245
+gpt-5.4-mini,non-latin,4,1941,1122,464,630,573,0.5315,0.5503,0.5256,0.5237,0.1316,0.3663,0.5377,555,170,281,251,463,0.5919,0.5517,0.5686,0.3515,0.2793,0.2345,0.5600
+harvey,en-latin,8,2099,783,816,259,1378,0.6517,0.5676,0.6683,0.1582,0.2207,0.2717,0.6138,411,118,144,168,1099,0.7784,0.6107,0.5897,0.1326,0.2140,0.2231,0.6000
+harvey,eu-latin,7,2081,747,692,382,1730,0.6767,0.5912,0.6483,0.1809,0.1966,0.2641,0.6184,685,116,134,146,967,0.8066,0.7326,0.7233,0.1312,0.1433,0.1448,0.7279
+harvey,non-latin,4,1161,585,1407,138,900,0.4918,0.3682,0.6162,0.1329,0.4462,0.3351,0.4610,468,65,350,77,564,0.6772,0.5300,0.7672,0.1201,0.3964,0.1220,0.6269
+harvey-2,en-latin,8,1922,584,1180,181,1446,0.6339,0.5214,0.7153,0.1112,0.3201,0.2330,0.6032,369,74,228,114,1147,0.7847,0.5499,0.6625,0.0904,0.3398,0.1670,0.6010
+harvey-2,eu-latin,7,1967,623,924,248,1859,0.6807,0.5598,0.6931,0.1177,0.2629,0.2405,0.6193,632,76,227,99,1010,0.8033,0.6759,0.7831,0.0893,0.2428,0.1073,0.7256
+harvey-2,non-latin,4,1171,556,1543,125,994,0.4933,0.3581,0.6323,0.1117,0.4719,0.3219,0.4572,461,75,381,41,638,0.6886,0.5027,0.7990,0.0604,0.4155,0.1399,0.6171
+legora-1,en-latin,8,1902,529,1269,124,1522,0.6405,0.5141,0.7444,0.0753,0.3430,0.2176,0.6082,296,22,355,45,1226,0.7829,0.4398,0.8154,0.0354,0.5275,0.0692,0.5714
+legora-1,eu-latin,7,2409,680,1422,225,2007,0.6549,0.5340,0.7269,0.1008,0.3152,0.2201,0.6157,838,64,299,89,1162,0.8157,0.6978,0.8456,0.0711,0.2490,0.0710,0.7646
+legora-1,non-latin,4,619,176,377,37,529,0.6605,0.5282,0.7440,0.0654,0.3217,0.2214,0.6178,210,3,85,25,309,0.8212,0.7047,0.8824,0.0749,0.2852,0.0141,0.7836
+legora-2,en-latin,8,1919,514,1267,130,1516,0.6425,0.5186,0.7487,0.0790,0.3424,0.2113,0.6128,330,46,297,60,1211,0.7927,0.4903,0.7569,0.0472,0.4413,0.1223,0.5951
+legora-2,eu-latin,7,2437,717,1381,235,2028,0.6568,0.5374,0.7191,0.1038,0.3045,0.2273,0.6151,850,94,264,91,1173,0.8184,0.7036,0.8213,0.0720,0.2185,0.0996,0.7579
+legora-2,non-latin,4,640,179,426,46,546,0.6456,0.5141,0.7399,0.0777,0.3422,0.2186,0.6066,229,5,88,23,323,0.8263,0.7112,0.8911,0.0665,0.2733,0.0214,0.7910

data/analysis/per_tradition.csv

@@ -0,0 +1,13 @@
+model,group,n_countries,tp,mismatch,missed,hallucinated,tn,accuracy,recall_when_filled,precision_when_emitted,hallucination_rate,miss_rate,wrong_when_both_filled,f1,cost_tp,cost_mismatch,cost_missed,cost_hallucinated,cost_tn,cost_accuracy,cost_recall_when_filled,cost_precision_when_emitted,cost_hallucination_rate,cost_miss_rate,cost_wrong_when_both_filled,cost_f1
+gemini/gemini-3.1-flash-lite,civil,11,4840,2242,1042,1151,2396,0.6200,0.5958,0.5879,0.3245,0.1283,0.3166,0.5918,1524,271,429,371,1649,0.7476,0.6853,0.7036,0.1837,0.1929,0.1510,0.6943
+gemini/gemini-3.1-flash-lite,common,8,1991,1246,456,423,1230,0.6025,0.5391,0.5440,0.2559,0.1235,0.3849,0.5415,398,119,155,201,1071,0.7557,0.5923,0.5543,0.1580,0.2307,0.2302,0.5727
+gpt-5.4-mini,civil,11,4916,2374,834,1621,1926,0.5862,0.6051,0.5517,0.4570,0.1027,0.3257,0.5772,1520,320,384,582,1438,0.6970,0.6835,0.6276,0.2881,0.1727,0.1739,0.6543
+gpt-5.4-mini,common,8,2018,1323,351,552,1091,0.5828,0.5466,0.5184,0.3360,0.0951,0.3960,0.5321,382,148,142,240,1028,0.7268,0.5685,0.4961,0.1893,0.2113,0.2792,0.5298
+harvey,civil,11,3242,1332,2099,520,2630,0.5978,0.4858,0.6364,0.1651,0.3146,0.2912,0.5510,1153,181,484,223,1531,0.7514,0.6342,0.7405,0.1271,0.2662,0.1357,0.6833
+harvey,common,8,2099,783,816,259,1378,0.6517,0.5676,0.6683,0.1582,0.2207,0.2717,0.6138,411,118,144,168,1099,0.7784,0.6107,0.5897,0.1326,0.2140,0.2231,0.6000
+harvey-2,civil,11,3138,1179,2467,373,2853,0.5985,0.4626,0.6691,0.1156,0.3636,0.2731,0.5470,1093,151,608,140,1648,0.7530,0.5902,0.7897,0.0783,0.3283,0.1214,0.6755
+harvey-2,common,8,1922,584,1180,181,1446,0.6339,0.5214,0.7153,0.1112,0.3201,0.2330,0.6032,369,74,228,114,1147,0.7847,0.5499,0.6625,0.0904,0.3398,0.1670,0.6010
+legora-1,civil,11,3028,856,1799,262,2536,0.6561,0.5328,0.7303,0.0936,0.3166,0.2204,0.6161,1048,67,384,114,1471,0.8168,0.6991,0.8527,0.0719,0.2562,0.0601,0.7683
+legora-1,common,8,1902,529,1269,124,1522,0.6405,0.5141,0.7444,0.0753,0.3430,0.2176,0.6082,296,22,355,45,1226,0.7829,0.4398,0.8154,0.0354,0.5275,0.0692,0.5714
+legora-2,civil,11,3077,896,1807,281,2574,0.6544,0.5324,0.7233,0.0984,0.3126,0.2255,0.6133,1079,99,352,114,1496,0.8201,0.7052,0.8351,0.0708,0.2301,0.0840,0.7647
+legora-2,common,8,1919,514,1267,130,1516,0.6425,0.5186,0.7487,0.0790,0.3424,0.2113,0.6128,330,46,297,60,1211,0.7927,0.4903,0.7569,0.0472,0.4413,0.1223,0.5951

data/analysis/quality/by_country.csv

@@ -0,0 +1,287 @@
+country,variable,n,empty,valid,invalid,pct_empty,pct_valid,pct_invalid
+al,court_cost_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+al,defendant_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+al,defendants_all_count,0,0,0,0,0.0000,0.0000,0.0000
+al,dispute_value_nominal,0,0,0,0,0.0000,0.0000,0.0000
+al,legal_subject_judgement,0,0,0,0,0.0000,0.0000,0.0000
+al,party_compensation_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+al,plaintiff_loosing_share,0,0,0,0,0.0000,0.0000,0.0000
+al,plaintiff_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+al,plaintiffs_all_count,0,0,0,0,0.0000,0.0000,0.0000
+al,trial_end_date,0,0,0,0,0.0000,0.0000,0.0000
+al,trial_start_date,0,0,0,0,0.0000,0.0000,0.0000
+am,court_cost_awarded_nominal,58,32,26,0,0.5517,0.4483,0.0000
+am,defendant_no1_ISIC1_industry_category,58,13,45,0,0.2241,0.7759,0.0000
+am,defendants_all_count,58,0,58,0,0.0000,1.0000,0.0000
+am,dispute_value_nominal,58,0,58,0,0.0000,1.0000,0.0000
+am,legal_subject_judgement,58,0,58,0,0.0000,1.0000,0.0000
+am,party_compensation_awarded_nominal,58,53,5,0,0.9138,0.0862,0.0000
+am,plaintiff_loosing_share,58,0,58,0,0.0000,1.0000,0.0000
+am,plaintiff_no1_ISIC1_industry_category,58,7,51,0,0.1207,0.8793,0.0000
+am,plaintiffs_all_count,58,0,58,0,0.0000,1.0000,0.0000
+am,trial_end_date,58,1,57,0,0.0172,0.9828,0.0000
+am,trial_start_date,58,17,41,0,0.2931,0.7069,0.0000
+at,court_cost_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+at,defendant_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+at,defendants_all_count,0,0,0,0,0.0000,0.0000,0.0000
+at,dispute_value_nominal,0,0,0,0,0.0000,0.0000,0.0000
+at,legal_subject_judgement,0,0,0,0,0.0000,0.0000,0.0000
+at,party_compensation_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+at,plaintiff_loosing_share,0,0,0,0,0.0000,0.0000,0.0000
+at,plaintiff_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+at,plaintiffs_all_count,0,0,0,0,0.0000,0.0000,0.0000
+at,trial_end_date,0,0,0,0,0.0000,0.0000,0.0000
+at,trial_start_date,0,0,0,0,0.0000,0.0000,0.0000
+au,court_cost_awarded_nominal,30,30,0,0,1.0000,0.0000,0.0000
+au,defendant_no1_ISIC1_industry_category,30,14,16,0,0.4667,0.5333,0.0000
+au,defendants_all_count,30,1,29,0,0.0333,0.9667,0.0000
+au,dispute_value_nominal,30,28,2,0,0.9333,0.0667,0.0000
+au,legal_subject_judgement,30,0,30,0,0.0000,1.0000,0.0000
+au,party_compensation_awarded_nominal,30,29,1,0,0.9667,0.0333,0.0000
+au,plaintiff_loosing_share,30,2,28,0,0.0667,0.9333,0.0000
+au,plaintiff_no1_ISIC1_industry_category,30,10,20,0,0.3333,0.6667,0.0000
+au,plaintiffs_all_count,30,0,30,0,0.0000,1.0000,0.0000
+au,trial_end_date,30,0,30,0,0.0000,1.0000,0.0000
+au,trial_start_date,30,23,7,0,0.7667,0.2333,0.0000
+be,court_cost_awarded_nominal,55,46,9,0,0.8364,0.1636,0.0000
+be,defendant_no1_ISIC1_industry_category,55,7,48,0,0.1273,0.8727,0.0000
+be,defendants_all_count,55,0,55,0,0.0000,1.0000,0.0000
+be,dispute_value_nominal,55,38,17,0,0.6909,0.3091,0.0000
+be,legal_subject_judgement,55,0,55,0,0.0000,1.0000,0.0000
+be,party_compensation_awarded_nominal,55,42,13,0,0.7636,0.2364,0.0000
+be,plaintiff_loosing_share,55,16,39,0,0.2909,0.7091,0.0000
+be,plaintiff_no1_ISIC1_industry_category,55,7,48,0,0.1273,0.8727,0.0000
+be,plaintiffs_all_count,55,0,55,0,0.0000,1.0000,0.0000
+be,trial_end_date,55,0,55,0,0.0000,1.0000,0.0000
+be,trial_start_date,55,39,16,0,0.7091,0.2909,0.0000
+br,court_cost_awarded_nominal,130,124,6,0,0.9538,0.0462,0.0000
+br,defendant_no1_ISIC1_industry_category,130,97,33,0,0.7462,0.2538,0.0000
+br,defendants_all_count,130,0,130,0,0.0000,1.0000,0.0000
+br,dispute_value_nominal,130,109,21,0,0.8385,0.1615,0.0000
+br,legal_subject_judgement,130,0,130,0,0.0000,1.0000,0.0000
+br,party_compensation_awarded_nominal,130,121,9,0,0.9308,0.0692,0.0000
+br,plaintiff_loosing_share,130,29,101,0,0.2231,0.7769,0.0000
+br,plaintiff_no1_ISIC1_industry_category,130,109,21,0,0.8385,0.1615,0.0000
+br,plaintiffs_all_count,130,0,130,0,0.0000,1.0000,0.0000
+br,trial_end_date,130,46,84,0,0.3538,0.6462,0.0000
+br,trial_start_date,130,102,28,0,0.7846,0.2154,0.0000
+ch,court_cost_awarded_nominal,130,0,130,0,0.0000,1.0000,0.0000
+ch,defendant_no1_ISIC1_industry_category,130,55,75,0,0.4231,0.5769,0.0000
+ch,defendants_all_count,130,0,130,0,0.0000,1.0000,0.0000
+ch,dispute_value_nominal,130,69,61,0,0.5308,0.4692,0.0000
+ch,legal_subject_judgement,130,0,130,0,0.0000,1.0000,0.0000
+ch,party_compensation_awarded_nominal,130,51,79,0,0.3923,0.6077,0.0000
+ch,plaintiff_loosing_share,130,2,128,0,0.0154,0.9846,0.0000
+ch,plaintiff_no1_ISIC1_industry_category,130,108,22,0,0.8308,0.1692,0.0000
+ch,plaintiffs_all_count,130,0,130,0,0.0000,1.0000,0.0000
+ch,trial_end_date,130,0,130,0,0.0000,1.0000,0.0000
+ch,trial_start_date,130,26,104,0,0.2000,0.8000,0.0000
+de,court_cost_awarded_nominal,130,125,5,0,0.9615,0.0385,0.0000
+de,defendant_no1_ISIC1_industry_category,130,66,64,0,0.5077,0.4923,0.0000
+de,defendants_all_count,130,23,107,0,0.1769,0.8231,0.0000
+de,dispute_value_nominal,130,35,95,0,0.2692,0.7308,0.0000
+de,legal_subject_judgement,130,0,130,0,0.0000,1.0000,0.0000
+de,party_compensation_awarded_nominal,130,130,0,0,1.0000,0.0000,0.0000
+de,plaintiff_loosing_share,130,49,81,0,0.3769,0.6231,0.0000
+de,plaintiff_no1_ISIC1_industry_category,130,97,33,0,0.7462,0.2538,0.0000
+de,plaintiffs_all_count,130,3,127,0,0.0231,0.9769,0.0000
+de,trial_end_date,130,0,130,0,0.0000,1.0000,0.0000
+de,trial_start_date,130,130,0,0,1.0000,0.0000,0.0000
+es,court_cost_awarded_nominal,130,78,52,0,0.6000,0.4000,0.0000
+es,defendant_no1_ISIC1_industry_category,130,9,121,0,0.0692,0.9308,0.0000
+es,defendants_all_count,130,3,127,0,0.0231,0.9769,0.0000
+es,dispute_value_nominal,130,31,99,0,0.2385,0.7615,0.0000
+es,legal_subject_judgement,130,1,129,0,0.0077,0.9923,0.0000
+es,party_compensation_awarded_nominal,130,72,58,0,0.5538,0.4462,0.0000
+es,plaintiff_loosing_share,130,16,114,0,0.1231,0.8769,0.0000
+es,plaintiff_no1_ISIC1_industry_category,130,13,117,0,0.1000,0.9000,0.0000
+es,plaintiffs_all_count,130,1,129,0,0.0077,0.9923,0.0000
+es,trial_end_date,130,4,125,1,0.0308,0.9615,0.0077
+es,trial_start_date,130,2,127,1,0.0154,0.9769,0.0077
+fr,court_cost_awarded_nominal,30,18,12,0,0.6000,0.4000,0.0000
+fr,defendant_no1_ISIC1_industry_category,30,5,25,0,0.1667,0.8333,0.0000
+fr,defendants_all_count,30,0,30,0,0.0000,1.0000,0.0000
+fr,dispute_value_nominal,30,17,13,0,0.5667,0.4333,0.0000
+fr,legal_subject_judgement,30,5,25,0,0.1667,0.8333,0.0000
+fr,party_compensation_awarded_nominal,30,30,0,0,1.0000,0.0000,0.0000
+fr,plaintiff_loosing_share,30,2,28,0,0.0667,0.9333,0.0000
+fr,plaintiff_no1_ISIC1_industry_category,30,4,26,0,0.1333,0.8667,0.0000
+fr,plaintiffs_all_count,30,0,30,0,0.0000,1.0000,0.0000
+fr,trial_end_date,30,0,30,0,0.0000,1.0000,0.0000
+fr,trial_start_date,30,30,0,0,1.0000,0.0000,0.0000
+ge,court_cost_awarded_nominal,112,27,85,0,0.2411,0.7589,0.0000
+ge,defendant_no1_ISIC1_industry_category,112,13,99,0,0.1161,0.8839,0.0000
+ge,defendants_all_count,112,0,112,0,0.0000,1.0000,0.0000
+ge,dispute_value_nominal,112,49,63,0,0.4375,0.5625,0.0000
+ge,legal_subject_judgement,112,0,112,0,0.0000,1.0000,0.0000
+ge,party_compensation_awarded_nominal,112,71,41,0,0.6339,0.3661,0.0000
+ge,plaintiff_loosing_share,112,25,87,0,0.2232,0.7768,0.0000
+ge,plaintiff_no1_ISIC1_industry_category,112,4,108,0,0.0357,0.9643,0.0000
+ge,plaintiffs_all_count,112,0,112,0,0.0000,1.0000,0.0000
+ge,trial_end_date,112,2,110,0,0.0179,0.9821,0.0000
+ge,trial_start_date,112,33,79,0,0.2946,0.7054,0.0000
+gh,court_cost_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+gh,defendant_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+gh,defendants_all_count,0,0,0,0,0.0000,0.0000,0.0000
+gh,dispute_value_nominal,0,0,0,0,0.0000,0.0000,0.0000
+gh,legal_subject_judgement,0,0,0,0,0.0000,0.0000,0.0000
+gh,party_compensation_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+gh,plaintiff_loosing_share,0,0,0,0,0.0000,0.0000,0.0000
+gh,plaintiff_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+gh,plaintiffs_all_count,0,0,0,0,0.0000,0.0000,0.0000
+gh,trial_end_date,0,0,0,0,0.0000,0.0000,0.0000
+gh,trial_start_date,0,0,0,0,0.0000,0.0000,0.0000
+hk,court_cost_awarded_nominal,10,10,0,0,1.0000,0.0000,0.0000
+hk,defendant_no1_ISIC1_industry_category,10,3,7,0,0.3000,0.7000,0.0000
+hk,defendants_all_count,10,0,10,0,0.0000,1.0000,0.0000
+hk,dispute_value_nominal,10,7,3,0,0.7000,0.3000,0.0000
+hk,legal_subject_judgement,10,0,10,0,0.0000,1.0000,0.0000
+hk,party_compensation_awarded_nominal,10,10,0,0,1.0000,0.0000,0.0000
+hk,plaintiff_loosing_share,10,1,9,0,0.1000,0.9000,0.0000
+hk,plaintiff_no1_ISIC1_industry_category,10,4,6,0,0.4000,0.6000,0.0000
+hk,plaintiffs_all_count,10,0,10,0,0.0000,1.0000,0.0000
+hk,trial_end_date,10,0,10,0,0.0000,1.0000,0.0000
+hk,trial_start_date,10,3,7,0,0.3000,0.7000,0.0000
+in,court_cost_awarded_nominal,24,24,0,0,1.0000,0.0000,0.0000
+in,defendant_no1_ISIC1_industry_category,24,4,20,0,0.1667,0.8333,0.0000
+in,defendants_all_count,24,5,19,0,0.2083,0.7917,0.0000
+in,dispute_value_nominal,24,7,17,0,0.2917,0.7083,0.0000
+in,legal_subject_judgement,24,0,24,0,0.0000,1.0000,0.0000
+in,party_compensation_awarded_nominal,24,18,6,0,0.7500,0.2500,0.0000
+in,plaintiff_loosing_share,24,1,23,0,0.0417,0.9583,0.0000
+in,plaintiff_no1_ISIC1_industry_category,24,11,13,0,0.4583,0.5417,0.0000
+in,plaintiffs_all_count,24,6,18,0,0.2500,0.7500,0.0000
+in,trial_end_date,24,0,24,0,0.0000,1.0000,0.0000
+in,trial_start_date,24,23,1,0,0.9583,0.0417,0.0000
+it,court_cost_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+it,defendant_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+it,defendants_all_count,0,0,0,0,0.0000,0.0000,0.0000
+it,dispute_value_nominal,0,0,0,0,0.0000,0.0000,0.0000
+it,legal_subject_judgement,0,0,0,0,0.0000,0.0000,0.0000
+it,party_compensation_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+it,plaintiff_loosing_share,0,0,0,0,0.0000,0.0000,0.0000
+it,plaintiff_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+it,plaintiffs_all_count,0,0,0,0,0.0000,0.0000,0.0000
+it,trial_end_date,0,0,0,0,0.0000,0.0000,0.0000
+it,trial_start_date,0,0,0,0,0.0000,0.0000,0.0000
+li,court_cost_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+li,defendant_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+li,defendants_all_count,0,0,0,0,0.0000,0.0000,0.0000
+li,dispute_value_nominal,0,0,0,0,0.0000,0.0000,0.0000
+li,legal_subject_judgement,0,0,0,0,0.0000,0.0000,0.0000
+li,party_compensation_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+li,plaintiff_loosing_share,0,0,0,0,0.0000,0.0000,0.0000
+li,plaintiff_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+li,plaintiffs_all_count,0,0,0,0,0.0000,0.0000,0.0000
+li,trial_end_date,0,0,0,0,0.0000,0.0000,0.0000
+li,trial_start_date,0,0,0,0,0.0000,0.0000,0.0000
+lu,court_cost_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+lu,defendant_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+lu,defendants_all_count,0,0,0,0,0.0000,0.0000,0.0000
+lu,dispute_value_nominal,0,0,0,0,0.0000,0.0000,0.0000
+lu,legal_subject_judgement,0,0,0,0,0.0000,0.0000,0.0000
+lu,party_compensation_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+lu,plaintiff_loosing_share,0,0,0,0,0.0000,0.0000,0.0000
+lu,plaintiff_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+lu,plaintiffs_all_count,0,0,0,0,0.0000,0.0000,0.0000
+lu,trial_end_date,0,0,0,0,0.0000,0.0000,0.0000
+lu,trial_start_date,0,0,0,0,0.0000,0.0000,0.0000
+np,court_cost_awarded_nominal,130,24,106,0,0.1846,0.8154,0.0000
+np,defendant_no1_ISIC1_industry_category,130,58,72,0,0.4462,0.5538,0.0000
+np,defendants_all_count,130,0,130,0,0.0000,1.0000,0.0000
+np,dispute_value_nominal,130,81,49,0,0.6231,0.3769,0.0000
+np,legal_subject_judgement,130,0,130,0,0.0000,1.0000,0.0000
+np,party_compensation_awarded_nominal,130,23,107,0,0.1769,0.8231,0.0000
+np,plaintiff_loosing_share,130,27,103,0,0.2077,0.7923,0.0000
+np,plaintiff_no1_ISIC1_industry_category,130,98,32,0,0.7538,0.2462,0.0000
+np,plaintiffs_all_count,130,0,130,0,0.0000,1.0000,0.0000
+np,trial_end_date,130,2,128,0,0.0154,0.9846,0.0000
+np,trial_start_date,130,55,75,0,0.4231,0.5769,0.0000
+nz,court_cost_awarded_nominal,29,29,0,0,1.0000,0.0000,0.0000
+nz,defendant_no1_ISIC1_industry_category,29,4,25,0,0.1379,0.8621,0.0000
+nz,defendants_all_count,29,0,29,0,0.0000,1.0000,0.0000
+nz,dispute_value_nominal,29,29,0,0,1.0000,0.0000,0.0000
+nz,legal_subject_judgement,29,0,29,0,0.0000,1.0000,0.0000
+nz,party_compensation_awarded_nominal,29,21,8,0,0.7241,0.2759,0.0000
+nz,plaintiff_loosing_share,29,0,29,0,0.0000,1.0000,0.0000
+nz,plaintiff_no1_ISIC1_industry_category,29,20,9,0,0.6897,0.3103,0.0000
+nz,plaintiffs_all_count,29,0,29,0,0.0000,1.0000,0.0000
+nz,trial_end_date,29,0,29,0,0.0000,1.0000,0.0000
+nz,trial_start_date,29,29,0,0,1.0000,0.0000,0.0000
+ph,court_cost_awarded_nominal,10,10,0,0,1.0000,0.0000,0.0000
+ph,defendant_no1_ISIC1_industry_category,10,4,6,0,0.4000,0.6000,0.0000
+ph,defendants_all_count,10,2,8,0,0.2000,0.8000,0.0000
+ph,dispute_value_nominal,10,3,7,0,0.3000,0.7000,0.0000
+ph,legal_subject_judgement,10,0,10,0,0.0000,1.0000,0.0000
+ph,party_compensation_awarded_nominal,10,8,2,0,0.8000,0.2000,0.0000
+ph,plaintiff_loosing_share,10,3,7,0,0.3000,0.7000,0.0000
+ph,plaintiff_no1_ISIC1_industry_category,10,4,6,0,0.4000,0.6000,0.0000
+ph,plaintiffs_all_count,10,1,9,0,0.1000,0.9000,0.0000
+ph,trial_end_date,10,0,10,0,0.0000,1.0000,0.0000
+ph,trial_start_date,10,10,0,0,1.0000,0.0000,0.0000
+rs,court_cost_awarded_nominal,26,26,0,0,1.0000,0.0000,0.0000
+rs,defendant_no1_ISIC1_industry_category,26,7,19,0,0.2692,0.7308,0.0000
+rs,defendants_all_count,26,0,26,0,0.0000,1.0000,0.0000
+rs,dispute_value_nominal,26,6,20,0,0.2308,0.7692,0.0000
+rs,legal_subject_judgement,26,0,26,0,0.0000,1.0000,0.0000
+rs,party_compensation_awarded_nominal,26,22,4,0,0.8462,0.1538,0.0000
+rs,plaintiff_loosing_share,26,2,24,0,0.0769,0.9231,0.0000
+rs,plaintiff_no1_ISIC1_industry_category,26,12,14,0,0.4615,0.5385,0.0000
+rs,plaintiffs_all_count,26,0,26,0,0.0000,1.0000,0.0000
+rs,trial_end_date,26,0,26,0,0.0000,1.0000,0.0000
+rs,trial_start_date,26,25,1,0,0.9615,0.0385,0.0000
+sg,court_cost_awarded_nominal,124,114,10,0,0.9194,0.0806,0.0000
+sg,defendant_no1_ISIC1_industry_category,124,17,107,0,0.1371,0.8629,0.0000
+sg,defendants_all_count,124,5,119,0,0.0403,0.9597,0.0000
+sg,dispute_value_nominal,124,75,49,0,0.6048,0.3952,0.0000
+sg,legal_subject_judgement,124,0,124,0,0.0000,1.0000,0.0000
+sg,party_compensation_awarded_nominal,124,112,12,0,0.9032,0.0968,0.0000
+sg,plaintiff_loosing_share,124,10,114,0,0.0806,0.9194,0.0000
+sg,plaintiff_no1_ISIC1_industry_category,124,34,90,0,0.2742,0.7258,0.0000
+sg,plaintiffs_all_count,124,4,120,0,0.0323,0.9677,0.0000
+sg,trial_end_date,124,4,120,0,0.0323,0.9677,0.0000
+sg,trial_start_date,124,10,114,0,0.0806,0.9194,0.0000
+tw,court_cost_awarded_nominal,130,70,60,0,0.5385,0.4615,0.0000
+tw,defendant_no1_ISIC1_industry_category,130,74,56,0,0.5692,0.4308,0.0000
+tw,defendants_all_count,130,3,127,0,0.0231,0.9769,0.0000
+tw,dispute_value_nominal,130,100,30,0,0.7692,0.2308,0.0000
+tw,legal_subject_judgement,130,0,130,0,0.0000,1.0000,0.0000
+tw,party_compensation_awarded_nominal,130,117,13,0,0.9000,0.1000,0.0000
+tw,plaintiff_loosing_share,130,15,115,0,0.1154,0.8846,0.0000
+tw,plaintiff_no1_ISIC1_industry_category,130,94,36,0,0.7231,0.2769,0.0000
+tw,plaintiffs_all_count,130,0,130,0,0.0000,1.0000,0.0000
+tw,trial_end_date,130,0,130,0,0.0000,1.0000,0.0000
+tw,trial_start_date,130,15,115,0,0.1154,0.8846,0.0000
+uk,court_cost_awarded_nominal,130,130,0,0,1.0000,0.0000,0.0000
+uk,defendant_no1_ISIC1_industry_category,130,8,122,0,0.0615,0.9385,0.0000
+uk,defendants_all_count,130,0,130,0,0.0000,1.0000,0.0000
+uk,dispute_value_nominal,130,110,20,0,0.8462,0.1538,0.0000
+uk,legal_subject_judgement,130,0,130,0,0.0000,1.0000,0.0000
+uk,party_compensation_awarded_nominal,130,124,6,0,0.9538,0.0462,0.0000
+uk,plaintiff_loosing_share,130,17,113,0,0.1308,0.8692,0.0000
+uk,plaintiff_no1_ISIC1_industry_category,130,13,117,0,0.1000,0.9000,0.0000
+uk,plaintiffs_all_count,130,0,130,0,0.0000,1.0000,0.0000
+uk,trial_end_date,130,0,130,0,0.0000,1.0000,0.0000
+uk,trial_start_date,130,47,83,0,0.3615,0.6385,0.0000
+us,court_cost_awarded_nominal,130,130,0,0,1.0000,0.0000,0.0000
+us,defendant_no1_ISIC1_industry_category,130,14,116,0,0.1077,0.8923,0.0000
+us,defendants_all_count,130,0,130,0,0.0000,1.0000,0.0000
+us,dispute_value_nominal,130,46,84,0,0.3538,0.6462,0.0000
+us,legal_subject_judgement,130,0,130,0,0.0000,1.0000,0.0000
+us,party_compensation_awarded_nominal,130,130,0,0,1.0000,0.0000,0.0000
+us,plaintiff_loosing_share,130,7,123,0,0.0538,0.9462,0.0000
+us,plaintiff_no1_ISIC1_industry_category,130,33,97,0,0.2538,0.7462,0.0000
+us,plaintiffs_all_count,130,0,130,0,0.0000,1.0000,0.0000
+us,trial_end_date,130,0,130,0,0.0000,1.0000,0.0000
+us,trial_start_date,130,11,119,0,0.0846,0.9154,0.0000
+xk,court_cost_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+xk,defendant_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+xk,defendants_all_count,0,0,0,0,0.0000,0.0000,0.0000
+xk,dispute_value_nominal,0,0,0,0,0.0000,0.0000,0.0000
+xk,legal_subject_judgement,0,0,0,0,0.0000,0.0000,0.0000
+xk,party_compensation_awarded_nominal,0,0,0,0,0.0000,0.0000,0.0000
+xk,plaintiff_loosing_share,0,0,0,0,0.0000,0.0000,0.0000
+xk,plaintiff_no1_ISIC1_industry_category,0,0,0,0,0.0000,0.0000,0.0000
+xk,plaintiffs_all_count,0,0,0,0,0.0000,0.0000,0.0000
+xk,trial_end_date,0,0,0,0,0.0000,0.0000,0.0000
+xk,trial_start_date,0,0,0,0,0.0000,0.0000,0.0000

data/analysis/quality/by_variable.csv

@@ -0,0 +1,12 @@
+variable,n,empty,valid,invalid,pct_empty,pct_valid,pct_invalid
+court_cost_awarded_nominal,1548,1047,501,0,0.6764,0.3236,0.0000
+defendant_no1_ISIC1_industry_category,1548,472,1076,0,0.3049,0.6951,0.0000
+defendants_all_count,1548,42,1506,0,0.0271,0.9729,0.0000
+dispute_value_nominal,1548,840,708,0,0.5426,0.4574,0.0000
+legal_subject_judgement,1548,6,1542,0,0.0039,0.9961,0.0000
+party_compensation_awarded_nominal,1548,1184,364,0,0.7649,0.2351,0.0000
+plaintiff_loosing_share,1548,224,1324,0,0.1447,0.8553,0.0000
+plaintiff_no1_ISIC1_industry_category,1548,682,866,0,0.4406,0.5594,0.0000
+plaintiffs_all_count,1548,15,1533,0,0.0097,0.9903,0.0000
+trial_end_date,1548,59,1488,1,0.0381,0.9612,0.0006
+trial_start_date,1548,630,917,1,0.4070,0.5924,0.0006

data/analysis/quant_results.tex

@@ -0,0 +1,21 @@
+% Auto-generated by legex-quant-results — do not edit by hand.
+% Aggregated over the 19 release jurisdictions; free-text legal_subject_judgement excluded from scoring (reported separately).
+\begin{table*}[t]
+\centering\small
+\begin{tabular}{@{}l rrrr@{\hskip 14pt} rrrr@{}}
+\toprule
+ & \multicolumn{4}{c}{10 structured fields} & \multicolumn{4}{c}{Cost block (4 fields)} \\
+\cmidrule(lr){2-5}\cmidrule(l){6-9}
+System & Recall & Precision & F1 & False fill & Recall & Precision & F1 & False fill \\
+\midrule
+Gemini & 65.9\,$\pm$\,0.5\% & 65.5\,$\pm$\,0.5\% & 0.66 & 30.2\,$\pm$\,0.6\% & \textbf{66.4\,$\pm$\,0.9\%} & 66.6\,$\pm$\,0.9\% & 0.67 & 17.4\,$\pm$\,0.7\% \\
+GPT-5.4-mini & \textbf{66.7\,$\pm$\,0.5\%} & 60.9\,$\pm$\,0.5\% & 0.64 & 41.8\,$\pm$\,0.7\% & 65.7\,$\pm$\,0.9\% & 59.6\,$\pm$\,0.9\% & 0.62 & 25.0\,$\pm$\,0.8\% \\
+Harvey & 58.8\,$\pm$\,0.5\% & 74.2\,$\pm$\,0.5\% & 0.66 & 16.2\,$\pm$\,0.5\% & 62.8\,$\pm$\,1.0\% & 69.4\,$\pm$\,1.0\% & 0.66 & 12.9\,$\pm$\,0.6\% \\
+Legora & 60.3\,$\pm$\,0.5\% & \textbf{82.6\,$\pm$\,0.5\%} & \textbf{0.70} & \textbf{8.7\,$\pm$\,0.4\%} & 61.9\,$\pm$\,1.0\% & \textbf{84.4\,$\pm$\,0.9\%} & \textbf{0.71} & \textbf{5.6\,$\pm$\,0.4\%} \\
+\bottomrule
+\end{tabular}
+\vskip 0.05in
+\caption{Headline extraction metrics over the 19 release jurisdictions (8 core and 11 preview), computed over each system's successfully processed cases (metric definitions in \cref{sec:systems}). The left block covers the ten structured fields, the right block the four cost-block variables; percentages carry $\pm$1\,SE. Denominators, in row order (Gemini/GPT-5.4-mini/Harvey/Legora), are $n_{\text{filled}}$\,=\,10276/10276/8999/8132 and $n_{\text{empty}}$\,=\,5194/5184/4781/4438 over the ten structured fields, and $n_{\text{filled}}$\,=\,2896/2896/2491/2172 and $n_{\text{empty}}$\,=\,3292/3288/3021/2856 over the cost block; Harvey and Legora have smaller denominators because of their ingest gaps. F1 standard errors are below 0.01 and omitted.
+The best value per column is marked in \textbf{bold} (lower is better for false fill).}
+\label{tab:overall}
+\end{table*}

data/analysis/tables/currency_frequencies.tex

@@ -0,0 +1,26 @@
+% Auto-generated by scripts/appendix_frequencies.py — do not edit by hand.
+\begin{table}[t]
+\caption{Currencies of the three monetary fields (\texttt{dispute\_value}, \texttt{court\_cost}, \texttt{party\_compensation}) over all filled cells of the release with a recorded currency (n\,=\,2,544). Labels as recorded by the annotators, with common synonyms mapped to ISO codes.}
+\label{tab:currencies}
+\vskip 0.05in
+\centering\small
+\begin{tabular}{@{}lrr@{}}
+\toprule
+\textbf{Currency} & \textbf{Cells} & \textbf{Share} \\
+\midrule
+EUR & 798 & 31.4\% \\
+BRL & 390 & 15.3\% \\
+TWD & 389 & 15.3\% \\
+NPR & 302 & 11.9\% \\
+CHF & 215 & 8.5\% \\
+GEL & 151 & 5.9\% \\
+INR & 72 & 2.8\% \\
+SGD & 58 & 2.3\% \\
+AMD & 52 & 2.0\% \\
+USD & 24 & 0.9\% \\
+GBP & 24 & 0.9\% \\
+Usd Equivalent In Gel & 12 & 0.5\% \\
+Other (26 labels) & 57 & 2.2\% \\
+\bottomrule
+\end{tabular}
+\end{table}

data/analysis/tables/diversity.tex

@@ -0,0 +1,32 @@
+% Auto-generated by scripts/diversity_stats.py — do not edit by hand.
+\begin{table}[t]
+\caption{Sample diversity along the annotated dimensions. \emph{Sectors} counts the distinct ISIC top-level sectors observed among plaintiffs and defendants (of 22 possible, A--V); \emph{multi-party} is the share of judgments with more than one plaintiff or defendant; the last two columns give the share of judgments with a coded dispute value and the fill rate over the four cost-block fields.}
+\label{tab:diversity}
+\vskip 0.05in
+\centering\small
+\begin{tabular}{@{}lrrrrr@{}}
+\toprule
+\textbf{Jurisdiction} & \textbf{$n$} & \textbf{Sectors} & \textbf{Multi-party} & \textbf{Dispute value} & \textbf{Cost block} \\
+\midrule
+Armenia & 58 & 13 & 21\% & 100\% & 63\% \\
+Australia & 30 & 9 & 70\% & 7\% & 26\% \\
+Belgium & 55 & 15 & 53\% & 31\% & 35\% \\
+Brazil & 130 & 11 & 14\% & 16\% & 26\% \\
+France & 25 & 7 & 52\% & 52\% & 47\% \\
+Georgia & 112 & 14 & 39\% & 56\% & 62\% \\
+Germany & 130 & 13 & 12\% & 73\% & 35\% \\
+Hong Kong & 10 & 5 & 10\% & 30\% & 30\% \\
+India & 24 & 9 & 50\% & 71\% & 48\% \\
+Nepal & 130 & 12 & 57\% & 38\% & 70\% \\
+New Zealand & 29 & 9 & 52\% & 0\% & 32\% \\
+Philippines & 10 & 4 & 40\% & 70\% & 40\% \\
+Serbia & 26 & 6 & 31\% & 77\% & 46\% \\
+Singapore & 124 & 17 & 33\% & 40\% & 37\% \\
+Spain & 129 & 11 & 54\% & 76\% & 62\% \\
+Switzerland & 130 & 13 & 29\% & 47\% & 77\% \\
+Taiwan & 130 & 13 & 25\% & 23\% & 42\% \\
+United Kingdom & 130 & 17 & 33\% & 15\% & 27\% \\
+United States & 130 & 17 & 58\% & 65\% & 40\% \\
+\bottomrule
+\end{tabular}
+\end{table}

data/analysis/tables/headline.tex

@@ -0,0 +1,191 @@
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Headline extraction metrics by jurisdiction for model \texttt{gemini/gemini-3.1-flash-lite}. Accuracy is per-cell. Recall is over the cells where the expert recorded a value. Hallucination rate is the share of legitimately-empty cells where the model invented a value. Cost-block $F_1$ aggregates over the four monetary variables.}
+\label{tab:headline-gemini-gemini-3-1-flash-lite}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Jurisdiction & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & Cost $F_1$ \\
+\midrule
+AM &  57.8\% &  50.3\% &  10.6\% & 0.432 \\
+AU &  67.3\% &  68.4\% &  34.3\% & 0.587 \\
+BE &  75.0\% &  67.8\% &   9.7\% & 0.648 \\
+BR &  62.7\% &  52.7\% &  28.0\% & 0.676 \\
+CH &  74.5\% &  79.3\% &  42.8\% & 0.906 \\
+DE &  58.3\% &  47.5\% &  29.0\% & 0.687 \\
+ES &  68.0\% &  67.5\% &  29.3\% & 0.751 \\
+FR &  61.5\% &  61.6\% &  38.7\% & 0.483 \\
+GE &  53.6\% &  49.0\% &  25.9\% & 0.450 \\
+HK &  65.5\% &  63.9\% &  31.6\% & 0.357 \\
+IN &  63.2\% &  63.7\% &  37.5\% & 0.640 \\
+NP &  47.1\% &  50.9\% &  64.1\% & 0.692 \\
+NZ &  73.0\% &  77.0\% &  32.6\% & 0.795 \\
+PH &  45.5\% &  36.9\% &  42.2\% & 0.421 \\
+RS &  69.2\% &  74.2\% &  40.0\% & 0.691 \\
+SG &  58.3\% &  51.4\% &  24.2\% & 0.510 \\
+TW &  63.7\% &  60.2\% &  29.5\% & 0.699 \\
+UK &  52.9\% &  42.9\% &  25.2\% & 0.525 \\
+US &  65.2\% &  58.6\% &  16.2\% & 0.627 \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Headline extraction metrics by jurisdiction for model \texttt{gpt-5.4-mini}. Accuracy is per-cell. Recall is over the cells where the expert recorded a value. Hallucination rate is the share of legitimately-empty cells where the model invented a value. Cost-block $F_1$ aggregates over the four monetary variables.}
+\label{tab:headline-gpt-5-4-mini}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Jurisdiction & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & Cost $F_1$ \\
+\midrule
+AM &  59.7\% &  53.8\% &  15.4\% & 0.469 \\
+AU &  60.0\% &  67.4\% &  50.4\% & 0.506 \\
+BE &  66.6\% &  63.7\% &  27.2\% & 0.494 \\
+BR &  55.5\% &  46.3\% &  36.0\% & 0.581 \\
+CH &  68.7\% &  82.1\% &  79.4\% & 0.855 \\
+DE &  55.5\% &  45.6\% &  33.0\% & 0.704 \\
+ES &  67.1\% &  69.4\% &  44.4\% & 0.720 \\
+FR &  59.1\% &  70.3\% &  63.1\% & 0.525 \\
+GE &  49.4\% &  49.1\% &  49.6\% & 0.348 \\
+HK &  64.5\% &  61.1\% &  28.9\% & 0.273 \\
+IN &  58.9\% &  70.1\% &  59.4\% & 0.693 \\
+NP &  45.4\% &  53.5\% &  78.0\% & 0.673 \\
+NZ &  62.7\% &  72.7\% &  51.5\% & 0.557 \\
+PH &  50.0\% &  49.2\% &  48.9\% & 0.667 \\
+RS &  70.3\% &  73.7\% &  36.0\% & 0.792 \\
+SG &  56.7\% &  51.6\% &  30.1\% & 0.442 \\
+TW &  61.3\% &  63.8\% &  43.6\% & 0.646 \\
+UK &  54.6\% &  47.0\% &  28.7\% & 0.498 \\
+US &  62.1\% &  56.7\% &  22.4\% & 0.589 \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Headline extraction metrics by jurisdiction for model \texttt{harvey}. Accuracy is per-cell. Recall is over the cells where the expert recorded a value. Hallucination rate is the share of legitimately-empty cells where the model invented a value. Cost-block $F_1$ aggregates over the four monetary variables.}
+\label{tab:headline-harvey}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Jurisdiction & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & Cost $F_1$ \\
+\midrule
+AM &  26.2\% &   8.3\% &   0.0\% & 0.420 \\
+AU &  73.6\% &  69.9\% &  21.2\% & 0.494 \\
+BE &  71.6\% &  62.0\% &   8.2\% & 0.614 \\
+BR &  59.8\% &  40.4\% &  22.0\% & 0.547 \\
+CH &  79.7\% &  78.5\% &  16.1\% & 0.862 \\
+DE &  62.6\% &  44.7\% &  16.1\% & 0.691 \\
+ES &  51.5\% &  47.9\% &  26.1\% & 0.756 \\
+FR &  62.7\% &  59.8\% &  31.5\% & 0.536 \\
+GE &  44.4\% &  35.6\% &  16.1\% & 0.478 \\
+HK &  65.5\% &  52.8\% &  10.5\% & 0.455 \\
+IN &  72.7\% &  68.5\% &  20.2\% & 0.615 \\
+NP &  43.2\% &  34.4\% &  26.5\% & 0.860 \\
+NZ &  79.6\% &  75.4\% &  14.4\% & 0.733 \\
+PH &  65.7\% &  61.9\% &  27.8\% & 0.541 \\
+RS &  78.9\% &  73.2\% &  10.4\% & 0.645 \\
+SG &  61.0\% &  50.6\% &  12.0\% & 0.508 \\
+TW &  67.1\% &  55.2\% &   9.8\% & 0.619 \\
+UK &  64.7\% &  58.6\% &  22.0\% & 0.559 \\
+US &  63.0\% &  53.2\% &   8.9\% & 0.709 \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Headline extraction metrics by jurisdiction for model \texttt{harvey-2}. Accuracy is per-cell. Recall is over the cells where the expert recorded a value. Hallucination rate is the share of legitimately-empty cells where the model invented a value. Cost-block $F_1$ aggregates over the four monetary variables.}
+\label{tab:headline-harvey-2}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Jurisdiction & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & Cost $F_1$ \\
+\midrule
+AM &  35.9\% &  20.8\% &   1.6\% & 0.420 \\
+AU &  76.1\% &  66.8\% &  10.9\% & 0.646 \\
+BE &  68.3\% &  54.6\% &   3.1\% & 0.609 \\
+BR &  62.8\% &  38.9\% &  14.8\% & 0.591 \\
+CH &  78.8\% &  75.9\% &  10.6\% & 0.855 \\
+DE &  64.4\% &  41.7\% &   8.8\% & 0.649 \\
+ES &  55.2\% &  50.7\% &  17.4\% & 0.750 \\
+FR &  62.7\% &  57.1\% &  26.1\% & 0.593 \\
+GE &  45.3\% &  35.5\% &  10.7\% & 0.449 \\
+HK &  58.0\% &  46.7\% &  17.9\% & 0.316 \\
+IN &  75.4\% &  66.1\% &   9.1\% & 0.818 \\
+NP &  39.9\% &  27.5\% &  24.6\% & 0.864 \\
+NZ &  79.9\% &  73.8\% &  11.4\% & 0.723 \\
+PH &  68.7\% &  66.7\% &  27.8\% & 0.686 \\
+RS &  71.6\% &  63.1\% &  12.5\% & 0.587 \\
+SG &  58.1\% &  46.3\% &  11.2\% & 0.439 \\
+TW &  65.9\% &  51.4\% &   5.9\% & 0.591 \\
+UK &  63.7\% &  53.4\% &  13.8\% & 0.643 \\
+US &  59.2\% &  47.1\% &   6.2\% & 0.607 \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Headline extraction metrics by jurisdiction for model \texttt{legora-1}. Accuracy is per-cell. Recall is over the cells where the expert recorded a value. Hallucination rate is the share of legitimately-empty cells where the model invented a value. Cost-block $F_1$ aggregates over the four monetary variables.}
+\label{tab:headline-legora-1}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Jurisdiction & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & Cost $F_1$ \\
+\midrule
+AM &   0.0\% &   0.0\% &   0.0\% & 0.000 \\
+AU &  74.2\% &  59.6\% &   5.1\% & 0.644 \\
+BE &  70.2\% &  57.3\% &   2.6\% & 0.677 \\
+BR &  64.7\% &  39.6\% &  11.7\% & 0.616 \\
+CH &  82.7\% &  80.0\% &   7.7\% & 0.903 \\
+DE &  51.6\% &  19.2\% &   8.4\% & 0.698 \\
+ES &  58.3\% &  52.3\% &  10.8\% & 0.748 \\
+FR &  63.3\% &  57.5\% &  25.2\% & 0.589 \\
+GE &   0.0\% &   0.0\% &   0.0\% & 0.000 \\
+HK &  69.1\% &  54.2\% &   2.6\% & 0.444 \\
+IN &  75.8\% &  67.3\% &  10.1\% & 0.805 \\
+NP &  43.2\% &  30.0\% &  17.9\% & 0.873 \\
+NZ &  81.5\% &  72.7\% &   6.1\% & 0.795 \\
+PH &  70.0\% &  66.2\% &  24.4\% & 0.600 \\
+RS &  75.2\% &  68.3\% &  12.0\% & 0.584 \\
+SG &  56.3\% &  41.5\% &   5.1\% & 0.352 \\
+TW &  71.0\% &  58.4\% &   4.7\% & 0.746 \\
+UK &  65.9\% &  54.5\% &   9.4\% & 0.615 \\
+US &  60.3\% &  48.8\% &   7.0\% & 0.577 \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Headline extraction metrics by jurisdiction for model \texttt{legora-2}. Accuracy is per-cell. Recall is over the cells where the expert recorded a value. Hallucination rate is the share of legitimately-empty cells where the model invented a value. Cost-block $F_1$ aggregates over the four monetary variables.}
+\label{tab:headline-legora-2}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Jurisdiction & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & Cost $F_1$ \\
+\midrule
+AM &   0.0\% &   0.0\% &   0.0\% & 0.000 \\
+AU &  72.1\% &  61.1\% &  12.4\% & 0.571 \\
+BE &  70.9\% &  58.0\% &   2.1\% & 0.702 \\
+BR &  64.0\% &  40.1\% &  13.3\% & 0.617 \\
+CH &  83.1\% &  80.5\% &   7.7\% & 0.890 \\
+DE &  52.6\% &  20.2\% &   8.4\% & 0.705 \\
+ES &  58.4\% &  52.3\% &  10.3\% & 0.729 \\
+FR &  63.0\% &  57.5\% &  26.1\% & 0.598 \\
+GE &   0.0\% &   0.0\% &   0.0\% & 0.000 \\
+HK &  63.6\% &  48.6\% &   7.9\% & 0.222 \\
+IN &  74.2\% &  65.5\% &  11.1\% & 0.732 \\
+NP &  43.5\% &  30.0\% &  17.3\% & 0.863 \\
+NZ &  83.4\% &  76.5\% &   6.8\% & 0.909 \\
+PH &  72.7\% &  69.2\% &  22.2\% & 0.688 \\
+RS &  76.6\% &  68.8\% &   9.0\% & 0.607 \\
+SG &  57.9\% &  43.6\% &   4.8\% & 0.413 \\
+TW &  70.6\% &  58.3\% &   5.7\% & 0.754 \\
+UK &  66.4\% &  54.9\% &   8.7\% & 0.645 \\
+US &  59.7\% &  47.7\% &   6.2\% & 0.596 \\
+\bottomrule
+\end{tabular}
+\end{table}

data/analysis/tables/per_field.tex

@@ -0,0 +1,143 @@
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Per-field extraction metrics, summed across jurisdictions, for model \texttt{gemini/gemini-3.1-flash-lite}.}
+\label{tab:per-field-gemini-gemini-3-1-flash-lite}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Variable & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & $F_1$ \\
+\midrule
+\texttt{court\_cost\_awarded\_nominal} &  85.5\% &  63.9\% &   4.1\% & 0.722 \\
+\texttt{defendant\_no1\_ISIC1\_industry\_category} &  50.7\% &  58.0\% &  65.9\% & 0.540 \\
+\texttt{defendants\_all\_count} &  75.1\% &  76.9\% &  90.5\% & 0.768 \\
+\texttt{dispute\_value\_nominal} &  56.2\% &  56.4\% &  44.0\% & 0.504 \\
+\texttt{legal\_subject\_judgement} &   3.6\% &   3.6\% &  83.3\% & 0.036 \\
+\texttt{party\_compensation\_awarded\_nominal} &  88.2\% &  71.4\% &   6.6\% & 0.727 \\
+\texttt{plaintiff\_loosing\_share} &  70.1\% &  71.3\% &  36.6\% & 0.727 \\
+\texttt{plaintiff\_no1\_ISIC1\_industry\_category} &  39.3\% &  54.1\% &  79.5\% & 0.429 \\
+\texttt{plaintiffs\_all\_count} &  87.9\% &  88.8\% & 100.0\% & 0.889 \\
+\texttt{trial\_end\_date} &  66.9\% &  67.6\% &  50.0\% & 0.680 \\
+\texttt{trial\_start\_date} &  52.4\% &  26.2\% &   9.7\% & 0.363 \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Per-field extraction metrics, summed across jurisdictions, for model \texttt{gpt-5.4-mini}.}
+\label{tab:per-field-gpt-5-4-mini}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Variable & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & $F_1$ \\
+\midrule
+\texttt{court\_cost\_awarded\_nominal} &  83.2\% &  65.5\% &   8.3\% & 0.690 \\
+\texttt{defendant\_no1\_ISIC1\_industry\_category} &  52.7\% &  64.7\% &  74.7\% & 0.578 \\
+\texttt{defendants\_all\_count} &  73.0\% &  73.8\% &  56.1\% & 0.769 \\
+\texttt{dispute\_value\_nominal} &  51.2\% &  57.1\% &  53.8\% & 0.486 \\
+\texttt{legal\_subject\_judgement} &   4.9\% &   4.9\% & 100.0\% & 0.049 \\
+\texttt{party\_compensation\_awarded\_nominal} &  81.9\% &  67.3\% &  13.6\% & 0.611 \\
+\texttt{plaintiff\_loosing\_share} &  66.3\% &  69.9\% &  55.2\% & 0.692 \\
+\texttt{plaintiff\_no1\_ISIC1\_industry\_category} &  38.7\% &  53.9\% &  80.5\% & 0.426 \\
+\texttt{plaintiffs\_all\_count} &  84.9\% &  85.4\% &  71.4\% & 0.875 \\
+\texttt{trial\_end\_date} &  66.5\% &  66.9\% &  44.1\% & 0.690 \\
+\texttt{trial\_start\_date} &  40.4\% &  41.6\% &  61.3\% & 0.364 \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Per-field extraction metrics, summed across jurisdictions, for model \texttt{harvey}.}
+\label{tab:per-field-harvey}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Variable & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & $F_1$ \\
+\midrule
+\texttt{court\_cost\_awarded\_nominal} &  85.6\% &  57.8\% &   1.9\% & 0.689 \\
+\texttt{defendant\_no1\_ISIC1\_industry\_category} &  53.6\% &  49.3\% &  37.3\% & 0.523 \\
+\texttt{defendants\_all\_count} &  71.0\% &  71.6\% &  48.6\% & 0.777 \\
+\texttt{dispute\_value\_nominal} &  58.5\% &  47.0\% &  32.9\% & 0.463 \\
+\texttt{legal\_subject\_judgement} &   3.9\% &   3.9\% &  83.3\% & 0.043 \\
+\texttt{party\_compensation\_awarded\_nominal} &  89.5\% &  66.9\% &   4.7\% & 0.717 \\
+\texttt{plaintiff\_loosing\_share} &  70.8\% &  71.4\% &  33.3\% & 0.740 \\
+\texttt{plaintiff\_no1\_ISIC1\_industry\_category} &  56.7\% &  41.8\% &  25.4\% & 0.452 \\
+\texttt{plaintiffs\_all\_count} &  74.4\% &  74.9\% &  83.3\% & 0.812 \\
+\texttt{trial\_end\_date} &  66.0\% &  65.0\% &  12.3\% & 0.761 \\
+\texttt{trial\_start\_date} &  48.5\% &  13.3\% &   4.4\% & 0.216 \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Per-field extraction metrics, summed across jurisdictions, for model \texttt{harvey-2}.}
+\label{tab:per-field-harvey-2}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Variable & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & $F_1$ \\
+\midrule
+\texttt{court\_cost\_awarded\_nominal} &  86.1\% &  57.0\% &   0.8\% & 0.679 \\
+\texttt{defendant\_no1\_ISIC1\_industry\_category} &  46.0\% &  33.6\% &  27.4\% & 0.406 \\
+\texttt{defendants\_all\_count} &  70.9\% &  71.8\% &  62.2\% & 0.777 \\
+\texttt{dispute\_value\_nominal} &  59.3\% &  33.1\% &  20.4\% & 0.387 \\
+\texttt{legal\_subject\_judgement} &   1.1\% &   0.7\% &   0.0\% & 0.009 \\
+\texttt{party\_compensation\_awarded\_nominal} &  88.4\% &  63.1\% &   4.9\% & 0.689 \\
+\texttt{plaintiff\_loosing\_share} &  71.8\% &  69.7\% &  15.9\% & 0.767 \\
+\texttt{plaintiff\_no1\_ISIC1\_industry\_category} &  54.2\% &  31.3\% &  18.8\% & 0.379 \\
+\texttt{plaintiffs\_all\_count} &  77.2\% &  77.6\% &  66.7\% & 0.832 \\
+\texttt{trial\_end\_date} &  65.8\% &  65.0\% &  14.0\% & 0.740 \\
+\texttt{trial\_start\_date} &  50.9\% &  15.1\% &   3.3\% & 0.247 \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Per-field extraction metrics, summed across jurisdictions, for model \texttt{legora-1}.}
+\label{tab:per-field-legora-1}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Variable & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & $F_1$ \\
+\midrule
+\texttt{court\_cost\_awarded\_nominal} &  92.0\% &  70.1\% &   0.9\% & 0.806 \\
+\texttt{defendant\_no1\_ISIC1\_industry\_category} &  62.2\% &  55.5\% &  22.5\% & 0.621 \\
+\texttt{defendants\_all\_count} &  66.5\% &  66.6\% &  37.1\% & 0.750 \\
+\texttt{dispute\_value\_nominal} &  65.0\% &  34.7\% &  12.3\% & 0.447 \\
+\texttt{legal\_subject\_judgement} &   2.2\% &   1.8\% &  16.7\% & 0.023 \\
+\texttt{party\_compensation\_awarded\_nominal} &  92.7\% &  79.1\% &   4.2\% & 0.797 \\
+\texttt{plaintiff\_loosing\_share} &  71.8\% &  69.3\% &  11.4\% & 0.784 \\
+\texttt{plaintiff\_no1\_ISIC1\_industry\_category} &  58.0\% &  36.8\% &  16.9\% & 0.438 \\
+\texttt{plaintiffs\_all\_count} &  70.5\% &  70.5\% &  28.6\% & 0.807 \\
+\texttt{trial\_end\_date} &  78.1\% &  77.8\% &  13.5\% & 0.825 \\
+\texttt{trial\_start\_date} &  56.0\% &  27.2\% &   3.6\% & 0.405 \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+% Auto-generated by legex-analysis.
+\begin{table}[h]
+\caption{Per-field extraction metrics, summed across jurisdictions, for model \texttt{legora-2}.}
+\label{tab:per-field-legora-2}
+\centering\small
+\begin{tabular}{@{}lrrrr@{}}
+\toprule
+Variable & Accuracy & Recall$_{\text{filled}}$ & Hallu. rate & $F_1$ \\
+\midrule
+\texttt{court\_cost\_awarded\_nominal} &  92.1\% &  70.4\% &   0.7\% & 0.808 \\
+\texttt{defendant\_no1\_ISIC1\_industry\_category} &  57.1\% &  50.7\% &  28.5\% & 0.570 \\
+\texttt{defendants\_all\_count} &  66.2\% &  65.8\% &  18.4\% & 0.755 \\
+\texttt{dispute\_value\_nominal} &  66.6\% &  39.7\% &  12.9\% & 0.488 \\
+\texttt{legal\_subject\_judgement} &   2.6\% &   2.2\% &  16.7\% & 0.028 \\
+\texttt{party\_compensation\_awarded\_nominal} &  92.5\% &  79.7\% &   4.5\% & 0.797 \\
+\texttt{plaintiff\_loosing\_share} &  72.5\% &  70.8\% &  16.3\% & 0.775 \\
+\texttt{plaintiff\_no1\_ISIC1\_industry\_category} &  59.2\% &  36.9\% &  14.8\% & 0.444 \\
+\texttt{plaintiffs\_all\_count} &  72.3\% &  72.4\% &  35.7\% & 0.816 \\
+\texttt{trial\_end\_date} &  77.4\% &  77.0\% &  13.2\% & 0.822 \\
+\texttt{trial\_start\_date} &  56.1\% &  27.2\% &   3.6\% & 0.406 \\
+\bottomrule
+\end{tabular}
+\end{table}

docs/Jurisdictions.md

@@ -0,0 +1,140 @@
+# Jurisdictions
+
+Research catalogue for scaling `legex` to new jurisdictions. Each row is a
+candidate country/court, what data source we'd use, how to isolate civil-law
+cases, and how hard the scraper looks. Two jurisdictions are already
+implemented; the others are prioritised into waves.
+
+Source links and contact names come from GitHub issue
+[#3](https://github.com/Engineers-for-Science-Initiative/legex/issues/3) and
+Daniel's compilation table. Link checks and source research were performed
+via WebFetch/WebSearch; HuggingFace datasets were preferred where available.
+
+## Status
+
+| #    | Country            | Court                        | Source type                                                                                       | Civil filter                                         | Dates                 | Recommendation                                                                           |
+| ---- | ------------------ | ---------------------------- | ------------------------------------------------------------------------------------------------- | ---------------------------------------------------- | --------------------- | ---------------------------------------------------------------------------------------- |
+| done | Switzerland        | Bundesgericht                | HF `voilaj/swiss-caselaw`                                                                       | chamber prefix 4* / 5*                               | 2015–2025            | Done                                                                                     |
+| done | France             | Cour de cassation            | PISTE Judilibre API                                                                               | chambers civ1/civ2/civ3                              | 2015–2025            | Done                                                                                     |
+| done | Australia          | High Court                   | HF `isaacus/high-court-of-australia-cases`                                                      | exclude "v The King/Queen"                           | 2015–2025            | Done (link via hcourt.gov.au per-case page, docket extracted from text)                  |
+| done | Germany            | BGH                          | HF `openlegaldata/court-decisions-germany` (gated)                                              | case-ID `<Roman> ZR/ZB/ZA`                         | 2015–2018            | Done (coverage tails off post-2018; gated T&C + HF_TOKEN)                                |
+| done | United States      | SCOTUS                       | HF `free-law/Caselaw_Access_Project` (gated)                                                    | exclude "United States v." + state-v heuristic       | 2015–2020            | Done (CAP dataset ends 2020; gated T&C + HF_TOKEN)                                       |
+| done | United Kingdom     | Supreme Court                | BAILII / supremecourt.uk<br />used https://nationalarchives.github.io/ds-find-caselaw-docs/public | all UKSC = civil/mixed                               | 2009–2025            | Blocked on BAILII (Anubis proof-of-work); supremecourt.uk feasible with paginated scrape |
+| done | Belgium            | Cour de Cassation            | juportal.be HTML                                                                                  | `C*` (civil) vs `P*` case-ID prefix              | 2015–2025            | Done (form POST + server-side pagination on juportal content pages)                      |
+| 3    | Austria            | OGH                          | RIS (ris.bka.gv.at) JSON API                                                                      | `Rechtsgebiete == Zivilrecht`                      | 2000–2025            | Wave 1 (API exists, but param schema needs more research)                                |
+| 4    | Armenia            | Cassation Court              | cassationcourt.am HTML                                                                            | dedicated `/precedent/civil-cases-advanced-search` | 2015–2025            | Wave 2 (CSRF-token form session handling required)                                       |
+| Done | New Zealand        | High Court + Court of Appeal | justice.govt.nz JDO HTML                                                                          | "Civil" dropdown                                     | 2015–2025            | Wave 2                                                                                   |
+| 6    | Hong Kong          | CFA + High Court             | legalref.judiciary.hk HTML                                                                        | court-level + case-type dropdowns                    | 1997–2025            | Wave 2                                                                                   |
+| 7    | Singapore          | Supreme Court + SICC         | judiciary.gov.sg HTML                                                                             | civil vs criminal case summaries                     | 2000–2025            | Wave 2                                                                                   |
+| 8    | Korea              | Supreme Court                | eng.scourt.go.kr HTML                                                                             | case designation `Da` (civil)                      | 2015–2025            | Wave 3 — Hangul                                                                         |
+| 9    | India              | Supreme Court                | sci.gov.in HTML + PDF                                                                             | free-text "civil appeal" search                      | 2013–2025            | Wave 3 — PDF heavy                                                                      |
+| 10   | Philippines        | Supreme Court                | elibrary.judiciary.gov.ph HTML                                                                    | monthly decision index                               | 1996–2025            | Wave 3                                                                                   |
+| 11   | Brazil             | STJ                          | scon.stj.jus.br HTML                                                                              | "Pesquisa Pronta" civil themes                       | 2015–2025            | Wave 3 — verify scope                                                                   |
+| 12   | Dominican Republic | Suprema Corte                | poderjudicial.gob.do HTML                                                                         | First Chamber (Civil + Commercial)                   | 2015–2025            | Wave 3                                                                                   |
+| 13   | Georgia            | Supreme Court                | supremecourt.ge HTML + AJAX                                                                       | chamber = "Civil"                                    | 2015–2025            | Wave 3 — needs Playwright                                                               |
+| 14   | Kazakhstan         | Supreme Court                | sud.gov.kz HTML                                                                                   | Bank of Judicial Acts by case type                   | 2009–2024            | Wave 3                                                                                   |
+| 15   | Ukraine            | Supreme Court                | reyestr.court.gov.ua HTML                                                                         | case form "Цивільні справи"            | 2015–2025            | Wave 3 — bot protection                                                                 |
+| 16   | Taiwan             | Judicial Yuan                | judgment.judicial.gov.tw ASPX                                                                     | `民事` case-type + ROC calendar                    | 2020–2025            | Wave 4 — ASPX + ROC calendar                                                            |
+| 17   | Nepal              | Supreme Court                | nkp.gov.np HTML                                                                                   | advanced-search case type                            | ~2015–2025           | Wave 4 — Bikram Sambat calendar, Nepali Unicode                                         |
+| 18   | Italy              | Corte di Cassazione          | italgiure.giustizia.it                                                                            | sezione civile (I–VI)                               | 1923–2025            | Blocked — subscription required                                                         |
+| 19   | Spain              | Tribunal Supremo             | poderjudicial.es                                                                                  | Sala de lo Civil                                     | 2000–2025            | Blocked — CGPJ bulk-use restriction, only 147 SC civil cases/yr                         |
+| 20   | China              | Supreme People's Court       | wenshu.court.gov.cn                                                                               | case-type filter                                     | 2015–2025            | Blocked — anti-bot + phone registration; check NYU Shanghai mirror                      |
+| 21   | Belarus            | Supreme Court                | court.gov.by                                                                                      | —                                                   | pre-2014 Plenums only | Blocked — no public decision database                                                   |
+
+Legend: Wave 1 = high confidence, should implement first. Wave 2 = moderate effort. Wave 3 = special challenges (encoding, bot protection, AJAX). Wave 4 = calendar/script conversion required. Blocked = needs a policy change, alternative source, or is not feasible.
+
+## Per-country detail
+
+Short notes per country with sample URL, rate/auth notes, and the scraper sketch. Implementation tickets will expand these.
+
+### Wave 1 — implement next
+
+- United Kingdom — Supreme Court (BAILII)
+
+  - Link: https://www.bailii.org/uk/cases/UKSC/
+  - Sample: https://www.bailii.org/uk/cases/UKSC/2024/24.html
+  - Civil filter: all UKSC decisions are civil/mixed; filter by year and case topic.
+  - Auth: none; free, no cookies, no anti-bot.
+  - Scraper: scrape `bailii.org/uk/cases/UKSC/<year>/` index pages; follow per-case HTML or PDF. ~100–150 civil cases/year 2015–2025. Stable structure.
+- Australia — High Court
+
+  - Link: https://www.hcourt.gov.au/cases-and-judgments/judgments
+  - HF dataset: `isaacus/high-court-of-australia-cases` (also `isaacus/open-australian-legal-corpus`).
+  - Civil filter: case-type classification; HCA mixes civil and constitutional.
+  - Scraper: reuse the HF parquet path, same pattern as CH. Trivial.
+- Austria — Oberster Gerichtshof (RIS)
+
+  - Link: https://www.ris.bka.gv.at
+  - Sample: https://www.ris.bka.gv.at/Dokumente/Justiz/JJT_20241024_OGH0002_0010OB00161_24V0000_000/
+  - Civil filter: senate codes like `Ob` for civil revision.
+  - Auth: none; German only.
+  - Scraper: query RIS by date+senate, paginate by case number.
+
+### Wave 2 — moderate effort
+
+- Germany — Bundesgerichtshof Zivilsenate
+
+  - Link: https://juris.bundesgerichtshof.de
+  - Civil filter: case prefix `ZR` (Zivilrecht) across senates I–XII.
+  - Auth: none (scraping BGH is legal per 2024 precedent).
+  - Scraper: parse `cgi-bin/rechtsprechung/document.py`; ~300 decisions/year civil.
+- Armenia — Court of Cassation
+
+  - Link: https://www.cassationcourt.am/en/precedent/civil-cases-advanced-search
+  - Civil filter: already has a dedicated civil-cases advanced-search form.
+  - Auth: none. Armenian / English / Russian UI.
+  - Scraper: POST the advanced-search form, parse results table.
+- United States — SCOTUS (via CourtListener)
+
+  - Official portal: https://www.supremecourt.gov/opinions/USReports.aspx (mixed civil/criminal, no filter).
+  - Preferred: CourtListener API (courtlistener.com/api) or HF "Pile of Law".
+  - Civil filter: apply post-hoc via case-type classification (SCOTUS cases are mostly certiorari; filter by case topic).
+  - Auth: CourtListener has free API keys with rate limits.
+  - Scraper: CourtListener API `/api/rest/v3/opinions/?court=scotus`.
+- New Zealand — High Court + Court of Appeal
+
+  - Link: https://www.justice.govt.nz/courts/decisions/jdo/
+  - Civil filter: "Civil" dropdown in JDO search.
+  - Auth: none. 3-day publication lag.
+  - Scraper: JDO search form, paginate, follow per-case HTML.
+- Hong Kong — Court of Final Appeal + High Court
+
+  - Link: https://legalref.judiciary.hk/lrs/common/ju/judgment.jsp
+  - Civil filter: court level + case type dropdowns (civil/criminal).
+  - Scraper: GET-based search form, fairly mechanical.
+- Singapore — Supreme Court + SICC
+
+  - Link: https://www.judiciary.gov.sg/judgments/judgments-case-summaries
+  - Civil filter: case summaries classified as civil vs criminal; SICC (commercial, 2015+) is a separate civil-only feed.
+  - Scraper: two feeds — general Supreme Court + SICC.
+
+### Wave 3 — special challenges
+
+- Korea: scourt.go.kr, case designation `Da` = civil appellate. Hangul full-text.
+- India: sci.gov.in, PDF-heavy; free-text "civil appeal" search. Multiple scripts possible.
+- Philippines: elibrary.judiciary.gov.ph, chronological monthly index, DataTables pagination.
+- Brazil: scon.stj.jus.br — "Pesquisa Pronta" is curated themes; verify whether the themed set is representative of civil cases or only a subset.
+- Dominican Republic: poderjudicial.gob.do — First Chamber (Civil + Commercial). Spanish.
+- Georgia: supremecourt.ge — AJAX modals; needs Playwright/Selenium rather than requests.
+- Kazakhstan: sud.gov.kz — Bank of Judicial Acts, filter by case type. Kazakh/Russian/English.
+- Ukraine: reyestr.court.gov.ua — case form "Цивільні справи". Anti-bot protection, "test mode" warnings; rate-limit gently.
+
+### Wave 4 — calendar / script conversion required
+
+- Taiwan: judgment.judicial.gov.tw is an ASP.NET form (stateful POST + cookies). Case-type filter `民事` (civil). Year input is ROC calendar (民國 N = Gregorian N+1911). Example: ROC 113 = 2024.
+- Nepal: nkp.gov.np uses Bikram Sambat calendar (~57 years ahead; 2015 AD ≈ 2072 BS). Nepali-Unicode-only search form. Calendar conversion helper required.
+
+### Blocked / investigate later
+
+- Italy — italgiure.giustizia.it is subscription-gated. Investigate whether the research group can obtain institutional access or negotiate bulk-export terms. No public bulk download.
+- Spain — poderjudicial.es CENDOJ prohibits bulk use (>100/day = ToS breach). Only ~147 Supreme Court civil cases since 2017 meet the Sala de lo Civil filter — insufficient for a 130-sample Goldenset unless we find alternative CENDOJ OpenData access.
+- China — wenshu.court.gov.cn requires Chinese mobile-phone registration and has aggressive anti-bot. Consider the NYU Shanghai Library snapshot or the CAIL2018 civil-cases dataset before attempting live scraping.
+- Belarus — court.gov.by only publishes Plenum explanations (non-binding jurisprudence) in the official gazette "Sudovy vestnik". No structured civil-decision database.
+
+## Key observations
+
+1. Seven jurisdictions are HuggingFace-ready or have usable public APIs (CH, FR already done; AU via `isaacus/high-court-of-australia-cases`; UK via BAILII; US via CourtListener). These are cheapest to implement and should be the next tickets.
+2. Most mid-tier countries are HTML scrapes with a dedicated civil filter at source (Austria, Germany, Belgium, Armenia, NZ, HK, Singapore, Dominican Republic). Each is ~1–2 days of scraper work.
+3. Four countries are effectively blocked without a policy change or alternative source (Italy, Spain, China, Belarus). Flag these to Adrian / Nerea / Jelle / contacts for guidance rather than burning engineering time.
+4. Non-Gregorian calendars and non-Latin scripts affect Taiwan (ROC), Nepal (Bikram Sambat), and the CJK countries. These need conversion utilities before scraping; consider a shared `legex/utils/calendar.py` when the first non-Gregorian jurisdiction ships.
+5. Prioritise "civil filter at source" over post-hoc filtering. Countries where the court already exposes a civil-only search (Armenia, Belgium C-prefix, Austria N Ob, Germany ZR) produce high-precision Goldensets with minimal noise.

goldensets/README.md

@@ -0,0 +1,309 @@
+---
+license: mit
+language:
+  - en
+  - de
+  - fr
+  - it
+  - es
+  - pt
+  - hy
+  - ka
+  - ne
+  - sr
+  - zh
+  - fil
+  - nl
+pretty_name: "LEGEX Goldensets: Expert-Coded Review-Table Annotations"
+size_categories:
+  - 1K<n<10K
+task_categories:
+  - text-classification
+  - token-classification
+  - text-generation
+tags:
+  - legal
+  - benchmark
+  - civil-judgments
+  - information-extraction
+  - llm-evaluation
+  - cross-jurisdictional
+dataset_info:
+  - config_name: default
+    features:
+      - name: case_id
+        dtype: string
+      - name: link
+        dtype: string
+      - name: full_text
+        dtype: string
+      - name: legal_subject_judgement
+        dtype: string
+      - name: trial_start_date
+        dtype: date32
+      - name: trial_end_date
+        dtype: date32
+      - name: dispute_value_nominal
+        dtype: string
+      - name: Currency_dispute_value_nominal
+        dtype: string
+      - name: plaintiff_loosing_share
+        dtype: float64
+      - name: court_cost_awarded_nominal
+        dtype: float64
+      - name: Currency_court_cost_awarded_nominal
+        dtype: string
+      - name: party_compensation_awarded_nominal
+        dtype: float64
+      - name: Currency_party_compensation_awarded_nominal
+        dtype: string
+      - name: plaintiffs_all_count
+        dtype: int64
+      - name: defendants_all_count
+        dtype: int64
+      - name: plaintiff_no1_ISIC1_industry_category
+        dtype: string
+      - name: defendant_no1_ISIC1_industry_category
+        dtype: string
+      - name: comment
+        dtype: string
+      - name: original_input
+        dtype: string
+      - name: annotator_id
+        dtype: string
+configs:
+  - config_name: default
+    data_files:
+      - split: armenia
+        path: data/am/goldenset_am.jsonl
+      - split: australia
+        path: data/au/goldenset_au.jsonl
+      - split: belgium
+        path: data/be/goldenset_be.jsonl
+      - split: brazil
+        path: data/br/goldenset_br.jsonl
+      - split: france
+        path: data/fr/goldenset_fr.jsonl
+      - split: georgia
+        path: data/ge/goldenset_ge.jsonl
+      - split: germany
+        path: data/de/goldenset_de.jsonl
+      - split: hong_kong
+        path: data/hk/goldenset_hk.jsonl
+      - split: india
+        path: data/in/goldenset_in.jsonl
+      - split: nepal
+        path: data/np/goldenset_np.jsonl
+      - split: new_zealand
+        path: data/nz/goldenset_nz.jsonl
+      - split: philippines
+        path: data/ph/goldenset_ph.jsonl
+      - split: serbia
+        path: data/rs/goldenset_rs.jsonl
+      - split: singapore
+        path: data/sg/goldenset_sg.jsonl
+      - split: spain
+        path: data/es/goldenset_es.jsonl
+      - split: switzerland
+        path: data/ch/goldenset_ch.jsonl
+      - split: taiwan
+        path: data/tw/goldenset_tw.jsonl
+      - split: united_kingdom
+        path: data/uk/goldenset_uk.jsonl
+      - split: united_states
+        path: data/us/goldenset_us.jsonl
+---
+
+# LEGEX Goldensets: Expert-Coded Review-Table Annotations
+
+This repository contains the expert-coded gold annotations for the LEGEX
+benchmark of civil-judgment review-table extraction. 1,548 judgments across
+19 jurisdictions have been annotated by hand against a shared 14-field schema
+covering monetary outcomes, cost allocation, party structure, and industry
+classification. Including independent secondary re-annotations, the release
+holds 1,974 annotation rows.
+
+## Dataset summary
+
+Legal review-table systems are increasingly used to extract structured facts
+from judgments, but there is little public evidence on their reliability in
+cross-jurisdictional legal settings. LEGEX is an expert-coded benchmark for
+civil-judgment review-table extraction. The current release contains an
+eight-jurisdiction core benchmark with at least 100 judgments per jurisdiction,
+each coded by two or more independent experts on a 28–30 case overlap and
+eleven preview jurisdictions for testing schema portability. The re-annotated
+overlap enables inter-annotator agreement (IAA) analysis on the core set.
+
+The word cloud below shows the topical spread of the free-text
+`legal_subject_judgement` labels over all annotated judgments (underscores
+stripped, stopwords and generic terms such as "law" removed):
+
+![Word cloud of the normalized legal-subject labels](assets/legal_subject_wordcloud.png)
+
+## Schema
+
+Each line in `data/<cc>/goldenset_<cc>.jsonl` is a JSON object with these
+keys:
+
+| Key | Type | Description                                                                                                                     |
+|-----|-----|---------------------------------------------------------------------------------------------------------------------------------|
+| `case_id` | string | Identifier within the source database.                                                                                          |
+| `link` | string | URL to the original judgment.                                                                                                   |
+| `full_text` | string \| null | Full judgment text used as model input.                                                                                         |
+| `legal_subject_judgement` | string | Short English subject of the case. Acts as the "this row has been substantively reviewed" marker.                               |
+| `trial_start_date` | YYYY-MM-DD \| null | Trial start date.                                                                                                               |
+| `trial_end_date` | YYYY-MM-DD \| null | Decision date.                                                                                                                  |
+| `dispute_value_nominal` | string \| null | Amount in dispute as a string (e.g. `"150000"`) or the literal `"nonpecuniary"`. |
+| `Currency_dispute_value_nominal` | string \| null | ISO-4217 currency code.                                                                                                         |
+| `plaintiff_loosing_share` | number [0, 1] \| null | Plaintiff's losing share.                                                                                                       |
+| `court_cost_awarded_nominal` | number \| null | Court fees awarded.                                                                                                             |
+| `Currency_court_cost_awarded_nominal` | string \| null | ISO-4217 currency code.                                                                                                         |
+| `party_compensation_awarded_nominal` | number \| null | Party compensation awarded.                                                                                                     |
+| `Currency_party_compensation_awarded_nominal` | string \| null | ISO-4217 currency code.                                                                                                         |
+| `plaintiffs_all_count` | integer \| null | Number of plaintiffs.                                                                                                           |
+| `defendants_all_count` | integer \| null | Number of defendants.                                                                                                           |
+| `plaintiff_no1_ISIC1_industry_category` | string \| null | ISIC Section A–U for the first plaintiff.                                                                                       |
+| `defendant_no1_ISIC1_industry_category` | string \| null | ISIC Section A–U for the first defendant.                                                                                       |
+
+A row is included when `legal_subject_judgement` is populated (the marker the
+annotator used to flag a row as substantively reviewed). Cells left empty by
+the annotator are stored as `null`. In the raw JSONL the trial dates are
+`YYYY-MM-DD` strings. The declared dataset features type them as `date32`, so
+the `datasets` library returns them as date objects.
+
+### Traceability fields
+
+Every record additionally carries provenance fields so each value is
+auditable:
+
+| Key | Type | Description                                                                                                                                                                                 |
+|-----|-----|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| `annotator_id` | string | Pseudonymous, salted-hash id of the coder, stable per person across jurisdictions.                                                                                                          |
+| `comment` | string \| null | Human-readable note listing every field changed during sanitization, e.g. `"The trial_end_date was sanitized from '2026_03_19' to '2026-03-19'."`, `null` when the record needed no change. |
+| `original_input` | string | JSON object mapping each changed field to its 1:1 pre-sanitization value, `"{}"` when nothing changed. This lets any consumer reconstruct the raw value.                                    |
+
+### Primary and secondary annotations
+
+Jurisdictions that received an independent re-annotation contain both the
+primary annotator's rows and the secondary annotators' rows, distinguished by
+`annotator_id`. Within each file all primary rows come first and re-annotation
+rows are appended after them, so **the first row per `case_id` is the primary
+gold annotation**. For benchmark evaluation, deduplicate to the primary rows.
+For inter-annotator agreement, group the duplicated `case_id`s by
+`annotator_id`.
+
+
+## Jurisdictions
+
+### Highest civil court per jurisdiction
+
+| Jurisdiction | Highest court (civil jurisdiction) | Language(s) |
+|-----|-----|-----|
+| Armenia | Court of Cassation (Civil Chamber) | Armenian |
+| Australia | High Court of Australia | English |
+| Belgium | Cour de cassation / Hof van Cassatie | French, Dutch |
+| Brazil | Superior Tribunal de Justiça | Portuguese |
+| France | Cour de cassation | French |
+| Georgia | Supreme Court of Georgia | Georgian |
+| Germany | Bundesgerichtshof | German |
+| Hong Kong | Court of Final Appeal | English, Chinese |
+| India | Supreme Court of India | English |
+| Nepal | Supreme Court of Nepal | Nepali |
+| New Zealand | Court of Appeal / High Court | English |
+| Philippines | Supreme Court of the Philippines | English, Filipino |
+| Serbia | Supreme Court of Cassation | Serbian |
+| Singapore | Supreme Court (incl. SICC) | English |
+| Spain | Tribunal Supremo (Sala de lo Civil) | Spanish |
+| Switzerland | Federal Supreme Court | German, French, Italian |
+| Taiwan | Supreme Court | Chinese |
+| United Kingdom | Supreme Court of the United Kingdom | English |
+| United States | Supreme Court of the United States | English |
+
+### Data sources, judgment counts and annotation rows
+
+`# judgments` counts unique expert-coded cases, `# rows` additionally counts
+independent secondary re-annotations of the same cases.
+
+| Jurisdiction | Data source | Time span | # judgments | # rows | Annotators |
+|-----|-----|-----|-----|-----|-----|
+| Armenia | [cassationcourt.am](https://www.cassationcourt.am) | 2024 – 2026 | 58 | 58 | 1 |
+| Australia | [HF isaacus/high-court-of-australia-cases](https://huggingface.co/datasets/isaacus/high-court-of-australia-cases) | 2015 – 2025 | 30 | 30 | 1 |
+| Belgium | [juportal.be](https://juportal.be) | 2015 – 2025 | 55 | 55 | 1 |
+| Brazil | [scon.stj.jus.br](https://scon.stj.jus.br) | 2002 – 2026 | 130 | 179 | 3 |
+| France | [Judilibre (PISTE) API](https://api.piste.gouv.fr/cassation/judilibre) | 2015 – 2025 | 30 | 30 | 1 |
+| Georgia | [supremecourt.ge](https://www.supremecourt.ge) | 2026 only | 112 | 161 | 3 |
+| Germany | [HF openlegaldata/court-decisions-germany](https://huggingface.co/datasets/openlegaldata/court-decisions-germany) | 2015 – 2022 | 130 | 180 | 3 |
+| Hong Kong | [legalref.judiciary.hk](https://legalref.judiciary.hk) | 2015 – 2026 | 10 | 10 | 1 |
+| India | [AWS Open Data Indian Supreme Court](https://registry.opendata.aws/indian-supreme-court-judgments) | 2017 – 2023 | 24 | 24 | 1 |
+| Nepal | [nkp.gov.np](https://nkp.gov.np) | 2021 – 2025 | 130 | 130 | 1 |
+| New Zealand | [justice.govt.nz JDO API](https://www.justice.govt.nz/jdo-search-api) | 2015 – 2025 | 29 | 29 | 1 |
+| Philippines | [elibrary.judiciary.gov.ph](https://elibrary.judiciary.gov.ph) | 2024 – 2025 | 10 | 10 | 1 |
+| Serbia | [vrh.sud.rs](https://vrh.sud.rs) | 2023 – 2025 | 26 | 26 | 1 |
+| Singapore | [sgcaselaw.com](https://sgcaselaw.com) | 2025 – 2026 | 124 | 172 | 3 |
+| Spain | [poderjudicial.es](https://www.poderjudicial.es) | 2025 – 2026 | 130 | 130 | 1 |
+| Switzerland | [HF voilaj/swiss-caselaw](https://huggingface.co/datasets/voilaj/swiss-caselaw) | 2024 – 2025 | 130 | 190 | 3 |
+| Taiwan | [judgment.judicial.gov.tw](https://judgment.judicial.gov.tw) | 2026 only | 130 | 180 | 3 |
+| United Kingdom | [caselaw.nationalarchives.gov.uk](https://caselaw.nationalarchives.gov.uk) | 2015 – 2025 | 130 | 190 | 3 |
+| United States | [HF free-law/Caselaw_Access_Project](https://huggingface.co/datasets/free-law/Caselaw_Access_Project) | 2020 – 2026 | 130 | 190 | 3 |
+
+Total: 1,548 expert-coded judgments, 1,974 annotation rows.
+
+Core benchmark (≥ 100 expert-coded judgments, coded by ≥ 2 independent
+annotators): Brazil, Georgia, Germany, Singapore, Switzerland, Taiwan,
+United Kingdom, United States. In each core jurisdiction 28–30 cases were
+independently re-annotated by one or two additional experts.
+Preview jurisdictions: the remaining eleven (Armenia, Australia, Belgium,
+France, Hong Kong, India, Nepal, New Zealand, Philippines, Serbia, Spain).
+Nepal reaches 130 cases but is single-annotated and therefore remains
+preview.
+
+## Loading
+
+```python
+from datasets import load_dataset
+
+# Single jurisdiction
+ds = load_dataset("legexbenchmark/goldensets", split="switzerland")
+ds = load_dataset("legexbenchmark/goldensets", split="united_states")
+
+# All jurisdictions
+from datasets import concatenate_datasets
+splits = [
+    "armenia", "australia", "belgium", "brazil", "france", "georgia",
+    "germany", "hong_kong", "india", "nepal", "new_zealand", "philippines",
+    "serbia", "singapore", "spain", "switzerland", "taiwan",
+    "united_kingdom", "united_states",
+]
+all_rows = concatenate_datasets([
+    load_dataset("legexbenchmark/goldensets", split=s) for s in splits
+])
+```
+
+## Limitations
+
+- Sample sizes: Nine of the nineteen jurisdictions have fewer than 100
+  expert-coded judgments; Nepal and Spain reach 100+ cases but are
+  single-annotated. These eleven are marked as preview and intended for
+  schema-portability checks rather than per-jurisdiction performance claims.
+- Time coverage: Source databases vary considerably (Georgia and Taiwan are
+  2026-only because earlier years were not freely available).
+- Schema portability: the 14 fields were designed against civil judgments
+  in common-law and Western European civil-law systems. Some fields
+  (e.g. `plaintiff_loosing_share`, ISIC categorisation) may not be a natural
+  fit for every jurisdiction.
+- Court selection: For Brazil, there are judgements that are not from the
+  highest possible court. This can be inferred with the case id. The Spain
+  goldenset draws mainly on the Tribunal Constitucional and the Tribunal
+  Económico-Administrativo Central rather than the Tribunal Supremo (Sala de
+  lo Civil); the court can likewise be inferred from the case_id.
+- Full-text quality: Text was extracted from heterogeneous sources (HTML,
+  PDF, API JSON). Layout artefacts and OCR errors are possible.
+
+## Citation
+
+Anonymous submission to the ICML 2026 AI for Law workshop. Citation block
+will be added after the camera-ready release.
+
+## License
+
+MIT.

goldensets/data/am/goldenset_am.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71224439e4fc6b668f91960f0ef4b5dd9ef9f3dd1f7ad858f004f0a9d833514b
+size 74216

goldensets/data/au/goldenset_au.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f49b54414f0de2776e819eb9bc1feddaa26e55219f5e0c0fc559499b2f6f18c
+size 908000

goldensets/data/be/goldenset_be.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce1c31d6a94b72a9aa9c7d3b55751edd4ca623136dbb5c9ea8eba9826ad49f6d
+size 641780

goldensets/data/br/goldenset_br.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0dce93837d96a913550d11179630284dde25203221caf771815c11fa5d7f9f9e
+size 2008499

goldensets/data/ch/goldenset_ch.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b05c92f82c0113dd18da050e105697c559c1046b6a963b076b3f21971bd9f84
+size 2770293

goldensets/data/de/goldenset_de.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95cbd2cd081e086c793b43a3c05ccf720b2630259bf2fcd6bff92926a1e18565
+size 3194110

goldensets/data/es/goldenset_es.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b027dd16fbc235c6731febf0771673289e36c81550bf2877dd1e50fe7c91c9fb
+size 3091836

goldensets/data/fr/goldenset_fr.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13dc24e2b6a432b1ffd0a79184797ab7c60ddba2dd8ee90ff35ca1e02cc955e2
+size 329994