yrdu/systemone-medbench

How Reliable Are System-One Decisions in Medicine?
A Comparative Study of Jev, Jev-like Models, and LLMs

Last updated · October 22, 2026   |   First report · October 13, 2026

↗ Interactive Results Website   ·   Paper on arXiv   ·   Dataset Overview   ·   Load the Data   ·   Source Licenses

SystemOne-MedBench is the Jev-format public subset of our medical decision reliability benchmark. It contains frozen decision inputs, separate gold references and identity mappings for paired evaluations. It supports research on performance, format robustness, abstention and probability-based selection.

Updates

Report first released on October 13, 2026. We continue to extend the evaluation to newly available open-weight and closed-source decision models.

Oct 22, 2026 · 🐘🦣 Completed evaluations of Elephant-1 (9B) and Mammoth-1 (27B), published SystemOne-MedBench on Hugging Face (10 public datasets, 4,408 parent entries and 29,582 requests), and prepared the evaluation toolkit for release.

Oct 20, 2026 · 🌐 Added PPLX-Decider v1.1 / v1.2 (27B) and the hosted Microsoft-Decision-1 and d1-ultra models.

Oct 17–18, 2026 · 📊 Extended evaluation to StartLux v1.1 Instant (0.8B / 4B / 9B / 27B) and D3-mini / D3-lite / D3-nano.

Oct 16, 2026 · 🧩 Added D3 (27B) under the same four-dimensional evaluation protocol.

Oct 15, 2026 · 🔄 Completed Jev 1.14 evaluation alongside the original Jev 1.13 baseline.

Oct 13, 2026 · 📄 Released the report, comparing Jev, representative Jev-like models, and general-purpose and medical LLMs across 15 datasets and four evaluation dimensions.

Dates track report publication and evaluation additions, not model release dates. The results website provides interactive comparisons; newly completed evaluations may appear there after a separate dashboard update.

Explore the results

Open Medical Decision Atlas → Explore interactive model comparisons across all four evaluation dimensions, with task-level results and metric explanations. The public website requires no login.

Website snapshot, October 22, 2026. The dashboard reports the study's evaluation results; this repository releases only the redistributable data subset. See cohort notes on the website.

Performance · Robustness · Abstention · Probability reliability

Release scope

Scope	Datasets	Parent entries	Evaluation requests
Public data in this repository	10	4,408	29,582
Complete paper benchmark	15	8,742	65,252

Public-subset scores are not full-paper scores. Five source-restricted or licensing-ambiguous subsets are withheld. See source rights and withheld data.

Only Jev-format text and structured numerical inputs are included. There are no LLM chat-format inputs, image tasks, sepsis experiments, agent tasks or thinking/routing experiments.

Contents

Task family	Dataset	Parent entries	Candidate classes	Label origin
A — Triage and Navigation	care_bench	135	4	Released CARE-Bench next-action reference for the retained conversation prefix.
A — Triage and Navigation	ramaswamy_triage	960	4	Author-released scenario triage reference; source-defined acceptable alternatives retained.
B — Clinical Judgment	medcalc_acute_assessment	41	2	Verified source calculator answer mapped to the stated decision threshold.
B — Clinical Judgment	medqa_decision	252	4	Original multiple-choice examination answer key.
B — Clinical Judgment	medxpertqa_acute	130	10	Original text-only expert multiple-choice answer key.
B — Clinical Judgment	sct_acute	78	5	Source expert-response distribution; maximum-score option(s) count as correct.
C — Safety Review	medec	509	2	Public MS source error/no-error annotation.
C — Safety Review	medval_review	388	4	Physician-assigned risk grade for source/target medical-text pair.
C — Safety Review	rxsafe	915	3	Source substitution-safety annotation, retaining source limitations.
D — Physiological Assessment	ctg_uci21	1000	3	Original UCI NSP fetal-state annotation, with all 21 numerical predictors.

Experimental dimensions

Variant	Public requests	Purpose
choice_letters	4,408	Original task performance
choice_numbers	4,408	Numeric relabeling (1–10 for ten candidates)
choice_shuffled	4,408	Fixed shuffled candidate order
binary_ovr	15,475	Independent Noul question for each candidate; OBS linear normalization
choice_none	382	Correct-answer replacement, eligible B datasets only
choice_abstain	501	Explicit optional abstention, B datasets only

Loading and file layout

The viewer contains gold-free records with a stable schema. body_json is a serialized native Jev request, not an LLM prompt. Parse it without sorting the candidate keys. Its request_sha256 is identical to the canonical request file. The viewer duplicates inference records for convenient browsing; it does not define an additional cohort.

import json
from datasets import load_dataset
rows = load_dataset("yrdu/systemone-medbench", "requests", split="test")
request = json.loads(rows[0]["body_json"])
# Send request["state"] and request["questions"] through the selected decision adapter.

For reproducible evaluation, pass an immutable commit hash from this dataset's Files and versions as revision. Download the repository and use data/jev/manifest.json with the companion evaluation runner for exact scoring.

data/jev/manifest.json
data/jev/<category>/<dataset>/answers.jsonl
data/jev/<category>/<dataset>/<variant>/jev_requests.jsonl
data/jev/<category>/<dataset>/<variant>/answers.jsonl
data/jev/<category>/<dataset>/<variant>/mappings.jsonl

Inference rows contain id, parent_id, body, and request_sha256. The body contains routing model, structured evidence state, and questions.decision. A question contains type, instructions, and ordered criteria. Choice criteria map displayed IDs to candidate meanings. Noul criteria give full affirmative and negative propositions. IDs and references are never model evidence.

The separately stored references contain a canonical label, an acceptable_labels set, or an SCT option_scores distribution. Mapping files align displayed labels to canonical decisions, or OBS option IDs to query IDs. Never score numeric/shuffled variants against unaligned original labels. NONE replacement is a changed task; ABSTAIN is withholding, not an extra correct diagnosis.

Construction and evaluation

This collection freezes the paper cohorts and transformations derived from the source benchmarks. Original inference payloads, detached labels and mappings are preserved byte-for-byte; selected subsets are exported without re-sampling. No new clinical adjudication is claimed. Each source dataset retains its own limitations. All six variants share parent IDs; they must not be treated as independent training/test samples.

Overall/category Accuracy and macro-F1, robustness and abstention results use equal dataset weighting. Probability-based selection pools the declared cohort and reports Accuracy/N/coverage at ≥0.70/0.75/0.80/0.85/0.90 and global top 5/10/15/20/25%. OBS uses q_i / sum(q) over affirmative probabilities, not another softmax. No selective-subset F1 is reported. Failed/overlength requests are not silently scored incorrect. Full details and executable scoring are in the companion code release.

Source rights and withheld data

See source-specific licenses and dataset_sources.json for provenance and notices. CARE and some other sources impose additional source or noncommercial terms. This collection is not blanket MIT/CC0 and must not be relicensed as such.

ESI (including its MIMIC component), MIMIC-IV-ED, ICD-proxy and MIMIC-CDM are not redistributed because they contain credentialed MIMIC-derived records. VitalDB-MAP is also withheld pending clarification of its disclosure terms. paper_scope.json contains aggregate counts and file hashes for reproducibility, not these records. Full-paper reproduction requires separately authorized access to the exact private frozen artifact; the public importer does not bypass source permissions.

Limitations and intended use

Research benchmarking only. Counts are dataset-qualified entries, not independent patients. Repeated scenarios, source annotation ambiguity, overlapping source cohorts, proxy targets and unknown model pretraining exposure can affect interpretation. Keep test references out of inference/training. These selective-accuracy results do not establish calibration, clinical safety, fairness or deployed routing efficiency. No attempt should be made to reidentify patients.

Complete input and detached-reference examples

The companion code contains one complete example per public dataset under examples/. The following is a real public CTG-21 example, not an invented model response. Its expected output is the detached reference, not a target probability value.

Expand the complete CTG-21 request and detached gold reference

{
  "model": "jev-1.13.0",
  "state": {
    "task": "Using the supplied numerical features extracted from a cardiotocography (CTG) recording, classify fetal state as normal, suspect, or pathologic. Select one class. This task does not ask whether delivery or another treatment should be performed.",
    "context": {
      "feature_descriptions": {
        "LB": "Fetal heart rate baseline (beats/min)",
        "AC": "Accelerations per second",
        "FM": "Fetal movements per second",
        "UC": "Uterine contractions per second",
        "DL": "Light decelerations per second",
        "DS": "Severe decelerations per second",
        "DP": "Prolonged decelerations per second",
        "ASTV": "Time with abnormal short-term variability (%)",
        "MSTV": "Mean short-term variability (source scale; no unit conversion)",
        "ALTV": "Time with abnormal long-term variability (%)",
        "MLTV": "Mean long-term variability (source scale; no unit conversion)",
        "Width": "FHR histogram width (beats/min)",
        "Min": "FHR histogram minimum (beats/min)",
        "Max": "FHR histogram maximum (beats/min)",
        "Nmax": "Number of histogram peaks",
        "Nzeros": "Number of histogram zeros",
        "Mode": "FHR histogram mode (beats/min)",
        "Mean": "FHR histogram mean (beats/min)",
        "Median": "FHR histogram median (beats/min)",
        "Variance": "FHR histogram variance (source scale; no unit conversion)",
        "Tendency": "Histogram tendency code: -1 left-skewed, 0 symmetric, 1 right-skewed (source coding)"
      },
      "features": {
        "LB": 140.0,
        "AC": 0.0,
        "FM": 0.0014184397163120568,
        "UC": 0.0028368794326241137,
        "DL": 0.0070921985815602835,
        "DS": 0.0,
        "DP": 0.0,
        "ASTV": 62.0,
        "MSTV": 1.2,
        "ALTV": 0.0,
        "MLTV": 3.9,
        "Width": 93.0,
        "Min": 66.0,
        "Max": 159.0,
        "Nmax": 5.0,
        "Nzeros": 2.0,
        "Mode": 142.0,
        "Mean": 136.0,
        "Median": 144.0,
        "Variance": 26.0,
        "Tendency": 1.0
      },
      "note": "Original 21 numerical features extracted from a CTG segment; not a waveform caption. Original spreadsheet precision retained; no unit conversion."
    },
    "options": {
      "A": "Normal fetal state",
      "B": "Suspect fetal state",
      "C": "Pathologic fetal state"
    },
    "task_definition": {
      "target": "Predict the expert-assigned fetal-state category for the supplied CTG recording: normal, suspect, or pathologic. These are mutually exclusive reference categories, not the ten-category morphological pattern labels.",
      "scope": "Input consists of extracted numerical summaries, not the original waveform. Do not invent unseen waveform features. No exact deterministic mapping from these summaries to the expert labels is supplied; this is classification using clinical knowledge, not execution of a fully specified waveform rule. Do not infer a delivery indication, treatment, or future neonatal outcome from the class label."
    },
    "field_descriptions": {
      "task": "Decision question and response instructions; not a patient observation.",
      "context": "Patient information available for this task; absent findings are unknown, not normal.",
      "options": "Map of option IDs to candidate meanings. IDs are response symbols, not measurements or severity scores.",
      "task_definition": "General endpoint, time-window and scope rules shared across cases; not the case outcome.",
      "context.features": "21 numerical features extracted from the same CTG segment, not a waveform or sequence of 21 timepoints.",
      "context.feature_descriptions": "Source feature names and short descriptions; explicit feature entries below include coding and unit caveats.",
      "context.note": "Representation and precision note; not a clinical observation.",
      "context.features.LB": "Baseline fetal heart rate (FHR), beats/min.",
      "context.features.AC": "FHR accelerations per second, not an event count or per-minute rate.",
      "context.features.FM": "Fetal movements per second.",
      "context.features.UC": "Uterine contractions per second.",
      "context.features.DL": "Light FHR decelerations per second, as extracted by the source.",
      "context.features.DS": "Severe FHR decelerations per second, as extracted by the source.",
      "context.features.DP": "Prolonged FHR decelerations per second, as extracted by the source.",
      "context.features.ASTV": "Percentage of time with abnormal short-term variability; percent, not a fraction.",
      "context.features.MSTV": "Mean short-term variability; original source scale, with no verified physical unit or conversion assumed.",
      "context.features.ALTV": "Percentage of time with abnormal long-term variability; percent, not a fraction.",
      "context.features.MLTV": "Mean long-term variability; original source scale, with no verified physical unit or conversion assumed.",
      "context.features.Width": "Width of the FHR histogram, in FHR-axis units (beats/min), not time duration.",
      "context.features.Min": "Lower FHR value of the histogram, beats/min; not the smallest histogram bin count.",
      "context.features.Max": "Upper FHR value of the histogram, beats/min; not the largest histogram bin count.",
      "context.features.Nmax": "Number of peaks in the FHR histogram, a count.",
      "context.features.Nzeros": "Number of histogram zeros as extracted by the source, a count; not episodes of absent heartbeat.",
      "context.features.Mode": "Mode of the FHR histogram, beats/min.",
      "context.features.Mean": "Mean of the FHR histogram, beats/min.",
      "context.features.Median": "Median of the FHR histogram, beats/min.",
      "context.features.Variance": "FHR histogram variance on the original source scale; no unverified physical-unit conversion assumed.",
      "context.features.Tendency": "Histogram asymmetry code: -1=left asymmetric, 0=symmetric, 1=right asymmetric. Not a worsening-risk score.",
      "field_descriptions": "Data dictionary: keys are input-relative paths; []=every array item; [0]/[1]=zero-based positional columns. Descriptions are not patient observations or answers.",
      "options.A": "Candidate meaning for response ID A; the option text defines the category/action, not whether it is correct.",
      "options.B": "Candidate meaning for response ID B; the option text defines the category/action, not whether it is correct.",
      "options.C": "Candidate meaning for response ID C; the option text defines the category/action, not whether it is correct.",
      "task_definition.target": "Definition of the decision target and candidate meanings.",
      "task_definition.scope": "Permitted evidence, missing-information rules and limits of the task.",
      "context.feature_descriptions.LB": "Explanatory text for LB; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.AC": "Explanatory text for AC; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.FM": "Explanatory text for FM; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.UC": "Explanatory text for UC; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.DL": "Explanatory text for DL; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.DS": "Explanatory text for DS; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.DP": "Explanatory text for DP; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.ASTV": "Explanatory text for ASTV; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.MSTV": "Explanatory text for MSTV; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.ALTV": "Explanatory text for ALTV; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.MLTV": "Explanatory text for MLTV; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.Width": "Explanatory text for Width; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.Min": "Explanatory text for Min; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.Max": "Explanatory text for Max; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.Nmax": "Explanatory text for Nmax; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.Nzeros": "Explanatory text for Nzeros; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.Mode": "Explanatory text for Mode; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.Mean": "Explanatory text for Mean; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.Median": "Explanatory text for Median; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.Variance": "Explanatory text for Variance; not a measurement. See the corresponding observation-field entry and general missingness/time rules.",
      "context.feature_descriptions.Tendency": "Explanatory text for Tendency; not a measurement. See the corresponding observation-field entry and general missingness/time rules."
    }
  },
  "questions": {
    "decision": {
      "type": "choice",
      "instructions": "Using the supplied numerical features extracted from a cardiotocography (CTG) recording, classify fetal state as normal, suspect, or pathologic. Select one class. This task does not ask whether delivery or another treatment should be performed.",
      "criteria": {
        "A": "Normal fetal state",
        "B": "Suspect fetal state",
        "C": "Pathologic fetal state"
      }
    }
  }
}

Detached reference:

{
  "id": "ctg:1826",
  "gold": {
    "label": "A"
  }
}

Citation

Please cite the paper when using the benchmark, and acknowledge the source datasets according to their respective licenses.

@article{du2026reliable,
  title={How Reliable Are System-One Decisions in Medicine? A Comparative Study of Jev, Jev-like Models, and LLMs},
  author={Du, Yanrui and Zhao, Danyang and Zhao, Sendong and Li, Wenhao and Gao, Yibo and Wang, Haochun and Xu, Zhuoer and Cui, Shiwen and Qin, Bing},
  journal={arXiv preprint arXiv:2610.18529},
  year={2026},
  doi={10.48550/arXiv.2610.18529}
}

---

Explore the evaluation website ↗ · Read the paper · Source-specific licenses