JusticeAxis-256

Data of JusticeAxis, a benchmark of 256 real-world criminal cases with audio, image and text evidence. A system reads a case and produces one judgment: the charge, the disposition, the sentence, and the reasoning.

• Paper: JusticeAxis: Benchmarking Legal Judgment between Rigid Rule Application and Ungrounded Discretion (arXiv:2610.00353)
• Project page: github.com/beita6969/JusticeAxis

This repository contains the case inputs and media (default config) and the reference answers (references config).

Why a reference-anchored task

The paper observes that existing methods swing between rigid statute matching and ungrounded discretion, while benchmarks score a label or a rubric, so how far a system sits between the two has not been compared quantitatively across models. JusticeAxis formalizes legal judgment as a reference-anchored task. Every task comes with three reference judgments, so a judgment can be located by which reference it lands closest to.

The task

The input is X = {X_m : m in M} with M a subset of {A, I, T, L}:

• A: audio,
• I: keyframes,
• T: text evidence together with the background of the case,
• L: the candidate statutes and the comparable precedents the charge turns on.

Missing modalities are allowed. Given X, a system generates a judgment J = (c, d, s, r): the charge, the disposition, the sentence, and the reasoning that cites what it relies on. The target judgment is withheld from solver inputs. A system must establish the source-supported facts and apply the relevant law; sentencing or appeal tasks may retain an earlier plea or conviction as task context.

Three references per case

Every task has three reference judgments of the same case:

1. the judgment as the court recorded it (recorded),
2. a rigid-rule judgment, which applies the matched statute to the established facts regardless of circumstances (rigid_rule),
3. an ungrounded-discretion judgment, which argues from the narrative on unsupported legal bases (ungrounded_discretion).

A system's judgment is compared with all three. How the comparison is made is defined under Metrics. The references are in data/references.jsonl, separate from the inputs.

Metrics

• Charge: exact-match accuracy and charge-family accuracy.

• Outcome: disposition accuracy and exact-sentence accuracy. Each is also reported conditioned on the coarser decision being right.

• Outcome distance. Sentences are compared by the log difference l(s, s') = |log(1+s) - log(1+s')|, and

  rho(J, J') = lambda_d * 1[d != d'] + lambda_s * l(s, s') / log(1 + s_max),

  where s_max is the longest sentence in the data.

• Placement. A judgment is placed at the reference it lands closest to under rho. The shares across the three references locate a system on the axis between rigid rule application and ungrounded discretion.

• Polarity. The signed share of the misses, Pol = (#rigid - #discretion) / #(not the recorded judgment), in [-1, 1]. +1 means reciting the statute, -1 means arguing from unsupported grounds.

Anchoring compares outcomes. How a judgment is argued is constrained rather than scored, so fluency earns no credit.

Results in the paper

The paper evaluates fourteen systems on 128 evaluation cases under one protocol: the evaluation package stays closed, retrieval runs over the provided X_L with no network lookup, every model receives the same X, and audio-less backbones receive its written description. Acc is the share of cases with the charge exactly correct, Sent. the share with the sentence exactly correct, Avg the mean of Acc, Fam, Disp and Sent. "Recorded" is the share of judgments landing closest to the recorded judgment. Values are copied from Table 2 of the paper, which also gives Wilson 95% intervals.

System	Acc	Sent.	Avg	Recorded (%)	Pol
Open-weight
Qwen3.8-27B	47.66	28.91	48.83	54.69	-0.345
Gemma 4 31B	32.81	17.97	32.03	34.38	-0.381
DeepSeek-V4-Flash (304B mixture-of-experts)	59.38	36.72	58.40	66.41	-0.302
Commercial
Claude Sonnet 5	78.12	67.97	78.52	84.38	0.500
Gemini 3.8 Flash	82.81	74.22	84.38	90.62	0.167
Grok 4.6	79.69	75.78	83.20	89.84	0.538
GPT-5.6	89.06	85.16	91.02	93.75	0.750
Peer frameworks on Qwen3.8-27B
Syllogism prompting	56.25	39.84	57.42	67.97	0.122
Syllogistic retrieval	67.19	55.47	68.75	77.34	0.241
Agentic legal search	67.97	57.81	71.48	80.47	0.840
Element agents	73.44	68.75	76.76	82.03	0.652
Courtroom simulation	64.06	58.59	70.12	78.12	0.500
Multimodal agents	74.22	64.06	74.80	78.91	0.259
JusticeAgent (Qwen3.8-27B)	78.91	70.31	80.86	85.16	0.263

What the paper reports:

• Failure turns direction with scale. The three open-weight backbones have negative Pol (arguing from grounds the record does not support). Every commercial model and every peer framework has positive Pol (returning the statutory default).
• The axis tracks competence. Placement and charge accuracy rank-correlate at 0.98 over the fourteen systems.
• The recording matters most. Replacing the audio by its written description costs 24.0 points of Avg, more than twice any other component in the ablation.

JusticeAgent

JusticeAgent is a harness around a frozen backbone. One element agent per element of the offence establishes the facts on a shared graph, the Element Graph, whose nodes are the elements required by the legal materials. A judge agent then applies the law in syllogistic form under contextual skills that carry experience of the circumstances. Skills are distilled from execution trajectories and admitted only under Bayesian credible bounds. On its own backbone it raises Acc from 47.66 to 78.91 and Pol from -0.345 to +0.263.

This release

Task rows	258 (130 original tasks from 128 cases, and 128 added cases)
Cases	256
Modalities	255 tasks with text, images and audio; 3 with text and images only
Images	1,441 files (919 PNG, 517 JPG, 5 JPEG)
Audio	435 files (311 WAV, 124 FLAC), no transcripts
Size	about 19.2 GB of media, almost all of it audio
Reference records	258 rows, each with the three references
Language	English inputs; the failure references and the summaries are in Chinese
Split	none

The paper reports results on 128 evaluation cases. This release does not say which cases those are; cohort only separates the original 128 cases from the 128 added later. Two baseline cases have two tasks each (JAVG-B02-008: federal and state; JAVG-B06-012: guilt and penalty), so there are 130 original task rows.

Layout

data/justiceaxis-256.jsonl            one row per task (258 rows): inputs
data/references.jsonl                 one row per task (258 rows): the three references
data/media/cases/<case_id>/I/*        images (keyframes)
data/media/cases/<case_id>/A/*        audio clips

Fields: default config (inputs)

Field	Type	Description
case_id	string	Case identifier, e.g. JAV200-002, NEW-UK-007.
record_id	string	Task identifier, <case_id>::<branch>. Join key to references.
branch_id	string	main, or the branch name for the two multi-task cases.
cohort	string	baseline128 (original) or new128 (added later).
language	string	Always en.
modalities	list of string	T text, I images, A audio.
event_group	string or null	Tasks that report the same event share a value. Group labels are retained only for the current selected tasks.
instruction	string	Task instruction shown to the solver.
requested_output	list of string	Output parts requested.
images, audio	list of {path}	Media paths relative to the repository root.
input_json	string (JSON)	The full solver input. Parse it with json.loads.

input_json has these keys: task (instruction, prediction boundary, requested output), modalities.T (case facts, context, legal context with candidate statutes, precedents and the sentencing framework, sentencing background), modalities.I (image descriptions), modalities.A (audio descriptions), evidence_alignment (links between facts, images and audio). The 130 original inputs and the 128 added inputs share one structure; where the material does not exist the field is empty. Participants are aliased (P1, P2, O1, V1, W1, ...).

Fields: references config

Field	Description
record_id, case_id, branch_id	As above.
available	True for all 258 tasks.
recorded	summary, answer_json (the full answer as JSON), outcome.
rigid_rule	answer, major_premise, minor_premise, conclusion, error_type, error_explanation, contrast_to_standard, citations_json, outcome.
ungrounded_discretion	answer, subjective_premise, factual_anchor, reasoning, conclusion, error_type, error_explanation, contrast_to_standard, citations_json, outcome.

outcome is the same structure for all three references: charge, disposition (one of custody_immediate, custody_suspended, community_order, fine, conditional_discharge_or_caution, convicted_no_sentence_reported, acquitted_or_dismissed, appeal_allowed, appeal_dismissed, charges_withdrawn, force_found_lawful, force_found_unlawful, outcome_not_reported, or null), component (always custodial_total), unit (always months), scope (whose sentence) and sentence (total custodial months; null when the source states no month or year figure, or for life terms; defendants are never added together and a driving ban is not added to custodial time).

Loading

import json
from datasets import load_dataset
from huggingface_hub import snapshot_download

inputs = load_dataset("beita6969/justiceaxis-256", "default", split="train")
refs = load_dataset("beita6969/justiceaxis-256", "references", split="train")
task = inputs[0]
reference = next(r for r in refs if r["record_id"] == task["record_id"])
print(task["record_id"], json.loads(task["input_json"])["task"]["instruction"][:80])
print(reference["available"], reference["recorded"]["outcome"] if reference["available"] else None)

# The media is about 19 GB. Download one task's files instead of everything:
paths = [p["path"] for p in task["images"]]          # add task["audio"] for audio
root = snapshot_download("beita6969/justiceaxis-256", repo_type="dataset", allow_patterns=paths)

Notes on the data

• The charge is not always withheld. In sentencing and appeal tasks the earlier conviction or plea is part of the input.
• Sentencing framework. sentencing_framework lists up to two Sentencing Council guideline pages (with a quote and the sha256 of the page text). Availability and source qualifications are recorded per input.
• Audio has no transcripts. Audio is shipped as the original clips with written descriptions and stated limits. Three tasks have no audio.
• Statutes are included in full in modalities.T.legal_context, which is why an input is large.

Scope and limitations

From the paper: the benchmark is retrospective and criminal only, and the harness is decision support, not a court.

Citation

@misc{tu2026justiceaxis,
  title         = {JusticeAxis: Benchmarking Legal Judgment between Rigid Rule Application and Ungrounded Discretion},
  author        = {Tu, Zhengkai and Zhang, Mingda and Wang, Zijia and Tang, Xiaoying and Huang, Jimmy},
  year          = {2026},
  eprint        = {2610.00353},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2610.00353}
}

License

Other (custom).