ClarusC64/clinical-intervention-sequencing-and-state-control-v0.2

Clinical Multi-Evidence State Integration Benchmark

CMESI v0.2

The Clinical Multi-Evidence State Integration Benchmark (CMESI) evaluates whether an AI system can reconstruct the evolving state of a complex clinical case across a sequence of heterogeneous evidence events.

CMESI does not test whether a model can identify a diagnosis from a static vignette alone. It tests whether the model can:

• maintain several competing clinical hypotheses simultaneously;
• distinguish diagnoses from investigations, treatments, referrals, constraints and monitoring processes;
• update only the states affected by each new evidence event;
• preserve unaffected states;
• revise earlier interpretations when evidence quality changes;
• reconstruct non-monotonic clinical trajectories;
• identify the evidence responsible for each transition;
• respect dependencies between diagnostic and operational states;
• distinguish decisive evidence from merely supportive evidence;
• and produce a complete structured representation of the final case state.

The benchmark is entirely synthetic. It contains no patient records, personal health information or identifiable clinical data.

---

Benchmark summary

Property	Value
Benchmark version	0.2
Total scenarios	48
Training scenarios	32
Test scenarios	16
Matched pairs	24
Training pairs	16
Test pairs	8
Language	English
Data format	JSON Lines
Data source	Synthetic
Patient data	None
Primary task	Structured clinical state reconstruction
License	MIT

Each matched pair shares a closely aligned prior context, tracked-item structure and evidence architecture while diverging at one or more decisive evidence points.

The paired design reduces the value of superficial keyword matching. A system must determine how small but causally important evidence differences alter the complete state trajectory.

---

What CMESI measures

Most clinical reasoning benchmarks collapse a case into a single answer:

What is the diagnosis?

CMESI instead asks:

What is the complete state of the case after every evidence event, and why did each tracked item change or remain unchanged?

A case may contain:

• several competing diagnoses;
• an investigation that becomes technically limited;
• a repeat test that supersedes an earlier result;
• a treatment that is planned, withheld, initiated, paused, resumed, switched or discontinued;
• a referral that is requested, scheduled, escalated, redirected or cancelled;
• a treatment-limiting constraint;
• a monitoring process;
• and downstream dependencies between these elements.

The benchmark therefore evaluates state integration, not isolated classification.

---

Core benchmark object

Each scenario is represented as an evolving multi-item state:

[ S_t = \left( s_t^{(1)}, s_t^{(2)}, \ldots, s_t^{(n)} \right) ]

where each (s_t^{(i)}) is the state of one tracked clinical item after evidence event (t).

Given:

• a prior state (S_0);
• an ordered evidence sequence (E_1,\ldots,E_T);
• item-specific state ontologies;
• workflow dependencies;
• and evidence-quality information;

the model must reconstruct:

[ S_0 \xrightarrow{E_1} S_1 \xrightarrow{E_2} \cdots \xrightarrow{E_T} S_T ]

The benchmark evaluates both the final state (S_T) and the intermediate transition path.

---

Key capabilities

1. Multi-item state tracking

A model must track multiple entities at once rather than reducing the case to one diagnosis.

Tracked items can include:

• diagnoses;
• investigations;
• treatments;
• referrals;
• constraints;
• adverse effects;
• monitoring processes.

2. Selective updating

An evidence event may affect one item, several items or no current item.

The model must avoid:

• changing every state after each event;
• erasing supported alternatives prematurely;
• confusing diagnostic evidence with workflow completion;
• or treating a treatment decision as direct diagnostic proof.

3. Non-monotonic reasoning

Clinical trajectories are not always monotonic.

A hypothesis may move through a sequence such as:

active_suspected
→ questioned
→ active_suspected
→ active_confirmed

An investigation may move through:

pending
→ technically_limited
→ repeat_required
→ performed
→ resolved

A treatment may move through:

active_confirmed
→ questioned
→ paused
→ resumed

CMESI explicitly scores whether these intermediate reversals are reconstructed correctly.

4. Evidence-quality integration

Evidence is not treated as equally reliable.

Cases include distinctions such as:

• technically adequate versus technically limited imaging;
• representative versus non-representative biopsy material;
• superficial versus deep microbiological samples;
• contaminated versus interpretable cultures;
• threshold versus strongly positive biomarkers;
• initial interpretation versus expert review;
• early imaging response versus later clinical resolution;
• valid versus treatment-contaminated dynamic testing.

A later expert review may not add a new disease finding directly. It may instead change how an earlier result should be weighted.

5. Cross-item dependencies

A diagnostic update may trigger operational changes elsewhere in the state.

Examples include:

diagnosis confirmed
→ treatment initiated
→ competing diagnosis ruled out
→ referral escalated

or:

diagnosis ruled out
→ treatment remains withheld
→ planned biopsy cancelled
→ referral redirected

These dependencies are explicitly represented and scored.

6. Workflow reasoning

CMESI distinguishes clinical findings from workflow state.

An investigation can be:

• pending;
• scheduled;
• performed;
• technically limited;
• inconclusive;
• repeat required;
• cancelled;
• resolved.

A result may change a diagnosis without completing the investigation pathway. Conversely, a workflow may resolve without changing every diagnostic state.

7. Decisive-evidence attribution

Models must identify which evidence events were decisive for each tracked item.

This is distinct from listing every relevant event.

For example:

• an early scan may support a diagnosis;
• an expert review may invalidate the scan interpretation;
• a repeat biopsy may become the decisive confirmation.

8. Minimal decisive subsets

The benchmark identifies one or more minimal evidence subsets sufficient to justify the final state of each item.

This tests whether a model can distinguish:

• the complete evidence history;
• decisive evidence;
• and the smallest sufficient combination of evidence.

9. Matched-pair discrimination

Each A/B pair contains deliberately similar cases with divergent outcomes.

A small evidence change may reverse:

• the confirmed diagnosis;
• which treatment is started;
• whether another treatment remains withheld;
• whether an investigation is repeated or cancelled;
• and which specialist pathway is activated.

---

Dataset structure

clinical-multi-evidence-state-integration-v0.2/
├── README.md
├── LICENSE
├── scorer.py
└── data/
    ├── train.jsonl
    └── test.jsonl

Splits

Split	Scenario range	Scenarios	Matched pairs
Train	CMESI-201-A to CMESI-216-B	32	16
Test	CMESI-217-A to CMESI-224-B	16	8

The test split contains complete gold annotations so that the benchmark can be evaluated locally and audited.

For blinded evaluations, organisers may distribute a derived test input file with the gold fields removed.

---

Clinical domains

CMESI v0.2 includes synthetic cases spanning multiple clinical domains, including:

• cardiology;
• dermatology;
• endocrinology;
• gastroenterology;
• haematology;
• hepatology;
• infectious disease;
• nephrology;
• neurology;
• oncology;
• ophthalmology;
• psychiatry;
• pulmonology;
• respiratory medicine;
• rheumatology;
• vascular medicine.

The benchmark is not intended to estimate real-world clinical performance within any individual speciality. The domains provide varied state-transition structures for evaluating evidence integration.

---

Record schema

Each line in a dataset file is one JSON object.

Identity and context

{
  "scenario_id": "CMESI-223-A",
  "pair_id": "CMESI-P223",
  "domain": "endocrinology",
  "prior_context": "..."
}

scenario_id

Unique scenario identifier.

The final A or B identifies the branch within a matched pair.

pair_id

Shared identifier linking the two matched cases.

domain

Clinical domain associated with the scenario.

prior_context

A concise account of the case before the evidence sequence begins.

---

Tracked items

{
  "tracked_items": [
    {
      "item_id": "P1",
      "label": "Primary adrenal insufficiency",
      "item_type": "diagnosis"
    },
    {
      "item_id": "I1",
      "label": "Short synacthen test pathway",
      "item_type": "investigation"
    },
    {
      "item_id": "T1",
      "label": "Glucocorticoid replacement therapy",
      "item_type": "treatment"
    }
  ]
}

Each tracked item contains:

Field	Meaning
item_id	Stable identifier within the scenario
label	Human-readable clinical description
item_type	State ontology assigned to the item

Common prefixes are used for readability:

Prefix	Typical item type
P	Diagnosis or clinical proposition
I	Investigation
T	Treatment
R	Referral or service pathway
C	Constraint or adverse effect
M	Monitoring process

The prefix is a convention rather than the formal source of the item type. Systems should use item_type.

---

Prior state

{
  "prior_state": {
    "P1": "active_suspected",
    "P2": "active_suspected",
    "I1": "pending",
    "T1": "withheld"
  }
}

The prior state defines the state of every tracked item before E1.

It is part of the reasoning problem and must not be inferred solely from the narrative.

---

Evidence sequence

{
  "evidence_sequence": [
    {
      "evidence_id": "E1",
      "sequence_index": 1,
      "evidence_type": "laboratory",
      "text": "Morning cortisol is low with elevated adrenocorticotropic hormone..."
    },
    {
      "evidence_id": "E2",
      "sequence_index": 2,
      "evidence_type": "treatment_event",
      "text": "Emergency hydrocortisone is administered..."
    }
  ]
}

Evidence must be interpreted in sequence.

Models should not treat the scenario as an unordered collection of findings because:

• later evidence may invalidate earlier evidence;
• workflow ordering may matter;
• a treatment may affect the validity of a later test;
• and an early transition may be reversed.

---

Gold final state

{
  "gold_final_state": {
    "P1": "active_confirmed",
    "P2": "ruled_out",
    "P3": "ruled_out",
    "I1": "resolved",
    "I2": "resolved",
    "T1": "initiated",
    "M1": "active_confirmed",
    "R1": "escalated"
  }
}

This is the state of every tracked item after the final evidence event.

Final-state accuracy is only one component of CMESI scoring.

---

Gold transition path

{
  "gold_transition_path": {
    "P1": [
      {
        "after_evidence_id": "E1",
        "state": "active_suspected",
        "transition": "no_change"
      },
      {
        "after_evidence_id": "E2",
        "state": "active_suspected",
        "transition": "no_change"
      },
      {
        "after_evidence_id": "E3",
        "state": "questioned",
        "transition": "question"
      }
    ]
  }
}

Every tracked item has one path entry after every evidence event.

Each path step records:

Field	Meaning
after_evidence_id	Evidence event just processed
state	Item state after the event
transition	Transition applied at the event

no_change is an explicit transition. It distinguishes correct preservation from missing output.

---

Final transition

{
  "gold_final_transition": {
    "P1": "confirm",
    "P2": "rule_out",
    "I1": "resolve",
    "T1": "initiate"
  }
}

The final transition is the principal transition that transforms the prior state into the final benchmark state.

It is not necessarily the transition attached to the final evidence event.

For example, a diagnosis may be confirmed at E5 and then remain unchanged through E8. Its gold_final_transition remains confirm.

---

Evidence roles

{
  "evidence_role_by_event": {
    "E1": {
      "P1": "supports",
      "P2": "weakens",
      "I1": "initiates_process",
      "T1": "supports"
    },
    "E2": {
      "P1": "supports",
      "P2": "insufficient",
      "I1": "limits_process",
      "T1": "initiates_process"
    }
  }
}

Evidence roles are item-specific.

The same event can:

• confirm one item;
• weaken another;
• initiate an investigation;
• justify withholding a treatment;
• and be irrelevant to another tracked item.

Common evidence roles include:

supports
confirms
reconfirms
weakens
contradicts
questions
excludes
supersedes
restores_support
initiates_process
advances_process
limits_process
requires_repeat
cancels_process
resolves_process
reinterprets_prior_evidence
supports_escalation
supports_deescalation
supports_withholding
supports_resumption
insufficient
irrelevant

---

Evidence quality

{
  "evidence_quality_by_event": {
    "E3": {
      "directness": "high",
      "technical_adequacy": "low",
      "specificity": "low",
      "source_reliability": "high",
      "temporal_relevance": "high"
    }
  }
}

Quality dimensions include:

Dimension	Question
directness	How directly does the event bear on the item or process?
technical_adequacy	Was the observation or procedure technically valid?
specificity	How specifically does it distinguish the relevant states?
source_reliability	How reliable is the evidence source?
temporal_relevance	How relevant is the evidence to the current state?

Values are:

low
moderate
high
not_applicable

Evidence-quality annotations describe the event itself. The item-specific effect remains encoded in evidence_role_by_event.

---

Item evidence traces

{
  "item_evidence_trace": {
    "P1": ["E1", "E2", "E3", "E4", "E5"],
    "T1": ["E1", "E2", "E5"]
  }
}

The trace identifies the evidence events relevant to each item.

A trace is broader than decisive evidence. It may contain supportive, contradictory, invalidated or operationally relevant events.

---

Decisive evidence

{
  "decisive_evidence": {
    "P1": ["E4", "E5"],
    "P2": ["E5"],
    "T1": ["E6"]
  }
}

decisive_evidence identifies events central to the final state or transition of each item.

An empty list is meaningful. It is used when an item is correctly preserved without requiring a new decisive event.

---

Minimal decisive subsets

{
  "minimal_decisive_subsets": {
    "P1": [
      ["E1", "E4", "E5"]
    ],
    "T1": [
      ["E5", "E6"]
    ],
    "P3": [
      []
    ]
  }
}

Each item maps to a collection of acceptable minimal evidence subsets.

A nested empty subset:

[[]]

means that no new evidence is required to justify preservation of the prior item state.

This differs from:

[]

which would indicate that no acceptable subset has been specified.

---

Changed and preserved items

{
  "changed_items": ["P1", "P2", "I1", "T1"],
  "preserved_items": ["P3"]
}

These fields test whether a model can distinguish genuine state updates from selective preservation.

Together they are intended to partition the tracked items.

---

Cross-item dependencies

{
  "cross_item_dependencies": [
    {
      "trigger_item": "P1",
      "trigger_state": "active_confirmed",
      "dependent_item": "T1",
      "required_transition": "initiate",
      "trigger_evidence_id": "E6"
    }
  ]
}

A dependency represents a required cross-item relationship.

It does not imply that the trigger state alone is universally sufficient in clinical practice. It specifies the dependency structure of the synthetic benchmark scenario.

Dependencies may represent:

• diagnostic confirmation triggering treatment;
• exclusion triggering treatment withholding;
• confirmation cancelling an unnecessary investigation;
• workflow completion enabling referral;
• or constraint resolution permitting treatment resumption.

---

Workflow constraints

{
  "workflow_constraints": [
    {
      "constraint_id": "W1",
      "type": "ordering",
      "item_id": "I1",
      "transition": "perform",
      "requires_prior_transition": "require_repeat"
    },
    {
      "constraint_id": "W2",
      "type": "prerequisite",
      "item_id": "T1",
      "transition": "initiate",
      "requires": {
        "item_id": "P1",
        "state": "active_confirmed"
      }
    }
  ]
}

Supported constraint structures include:

• ordering;
• prerequisite;
• dependency;
• cancellation condition;
• escalation condition;
• resumption condition;
• scheduling condition;
• mutual exclusion.

The official scorer evaluates whether the predicted path and final state satisfy these constraints.

---

Valid alternative paths

{
  "valid_alternative_paths": [
    {
      "path_id": "ALT-1",
      "applies_to_item": "P1",
      "allowed_final_state": "active_suspected",
      "allowed_transition_path": [
        {
          "after_evidence_id": "E1",
          "state": "active_suspected",
          "transition": "no_change"
        }
      ],
      "rationale": "The evidence permits a more cautious calibrated state."
    }
  ]
}

Most scenarios have one canonical path and no alternatives.

A small number of deliberately ambiguous scenarios include valid alternative paths. These are used when more than one calibration is defensible from the synthetic evidence.

Alternative paths:

• apply only to the explicitly named item;
• do not replace the canonical annotations for other items;
• and are accepted by the official scorer.

---

Integration labels

{
  "integration_types": [
    "TECHNICAL_LIMITATION",
    "PROCESS_REPETITION",
    "TREATMENT_ACTIVATION",
    "CROSS_ITEM_CASCADE"
  ],
  "reasoning_patterns": [
    "QUALITY_WEIGHTED_REINTERPRETATION",
    "REPEAT_IMAGING_CONFIRMATION",
    "DEPENDENT_TREATMENT_UPDATE"
  ]
}

These fields support mechanism-specific evaluation slices.

They are descriptive benchmark annotations rather than model output labels required for the primary task.

---

Complexity annotations

{
  "mechanism_complexity": 7,
  "complexity_components": {
    "tracked_item_count": 8,
    "evidence_event_count": 8,
    "competing_hypothesis_count": 3,
    "contradictory_evidence_count": 2,
    "non_monotonic_transition_count": 6,
    "cross_item_dependency_count": 4,
    "workflow_transition_count": 12,
    "historical_reinterpretation_count": 1
  }
}

mechanism_complexity is an ordinal benchmark-design label from 1 to 7.

It is not a validated measure of clinical difficulty.

The component fields make the basis of the assigned complexity more transparent.

---

Annotation confidence

{
  "annotation_confidence": "high",
  "confidence_rationale": "The technically invalid first test is superseded by adequate repeat testing..."
}

Possible values are:

high
medium
low

Confidence refers to the determinacy of the synthetic benchmark annotation, not to confidence in a real patient diagnosis.

---

Item-state ontologies

Diagnosis states

inactive
active_suspected
active_confirmed
downgraded
questioned
ruled_out
resolved
superseded
historical

Investigation states

inactive
not_requested
requested
pending
scheduled
performed
technically_limited
inconclusive
repeat_required
repeated
cancelled
resolved
superseded

Treatment states

inactive
planned
questioned
withheld
initiated
active_confirmed
paused
resumed
switched
escalated
deescalated
completed
discontinued
historical
resolved

Referral states

inactive
requested
pending
scheduled
active_confirmed
escalated
deescalated
switched
cancelled
resolved

Constraint and adverse-effect states

inactive
active_suspected
active_confirmed
questioned
downgraded
resolved
superseded
historical

Monitoring-process states

inactive
pending
requested
scheduled
performed
active_suspected
active_confirmed
questioned
resolved
cancelled
historical

---

Transition ontology

no_change
suspect
confirm
reconfirm
question
downgrade
rule_out
reactivate
resolve
supersede
request
schedule
perform
mark_limited
mark_inconclusive
require_repeat
repeat
cancel
plan
withhold
initiate
pause
resume
switch
escalate
deescalate
complete
discontinue
mark_historical
deactivate

States and transitions are separate.

For example:

state: active_confirmed
transition: confirm

means the item has entered the confirmed state.

Later:

state: active_confirmed
transition: reconfirm

means new evidence has reinforced the state.

Later still:

state: active_confirmed
transition: no_change

means the evidence did not alter the item state.

---

Prediction format

Predictions must be supplied as JSON Lines with one object per scenario.

Required structure:

{
  "scenario_id": "CMESI-223-A",
  "final_state": {
    "P1": "active_confirmed",
    "P2": "ruled_out"
  },
  "transition_path": {
    "P1": [
      {
        "after_evidence_id": "E1",
        "state": "active_suspected",
        "transition": "no_change"
      }
    ]
  },
  "final_transition": {
    "P1": "confirm",
    "P2": "rule_out"
  },
  "evidence_role_by_event": {
    "E1": {
      "P1": "supports",
      "P2": "weakens"
    }
  },
  "decisive_evidence": {
    "P1": ["E5", "E6"],
    "P2": ["E5"]
  },
  "minimal_decisive_subsets": {
    "P1": [
      ["E1", "E5", "E6"]
    ],
    "P2": [
      ["E5"]
    ]
  },
  "changed_items": ["P1", "P2"],
  "preserved_items": []
}

A prediction should include:

• every tracked item in final_state;
• one transition-path entry per evidence event for every item;
• every tracked item in final_transition;
• an item-specific evidence role for each event;
• decisive evidence for every item;
• minimal decisive subsets for every item;
• changed items;
• preserved items.

Missing values are scored as incorrect rather than inferred by the scorer.

---

Generating a prediction template

The scorer can generate a blank prediction file with the required structure:

python scorer.py \
  --gold data/test.jsonl \
  --write-template predictions/test_template.jsonl

The generated template contains:

• all scenario identifiers;
• all tracked item identifiers;
• all evidence identifiers;
• blank state and transition fields;
• and the required output containers.

---

Loading the dataset

Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset(
    "YOUR_USERNAME/clinical-multi-evidence-state-integration-v0.2"
)

print(dataset)
print(dataset["train"][0])

Local JSON loading

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files={
        "train": "data/train.jsonl",
        "test": "data/test.jsonl",
    },
)

Standard Python

import json
from pathlib import Path


def read_jsonl(path: str) -> list[dict]:
    records = []

    with Path(path).open("r", encoding="utf-8") as handle:
        for line in handle:
            line = line.strip()

            if line:
                records.append(json.loads(line))

    return records


train_records = read_jsonl("data/train.jsonl")
test_records = read_jsonl("data/test.jsonl")

print(len(train_records))
print(len(test_records))

---

Gold-file validation

Validate the training split:

python scorer.py \
  --gold data/train.jsonl \
  --validate-only

Validate the test split:

python scorer.py \
  --gold data/test.jsonl \
  --validate-only

Use strict validation to treat warnings as failures:

python scorer.py \
  --gold data/test.jsonl \
  --validate-only \
  --strict-validation

---

Running an evaluation

python scorer.py \
  --gold data/test.jsonl \
  --predictions predictions.jsonl \
  --output results/evaluation.json

Suppress the console report:

python scorer.py \
  --gold data/test.jsonl \
  --predictions predictions.jsonl \
  --output results/evaluation.json \
  --quiet

Omit per-scenario details from the JSON report:

python scorer.py \
  --gold data/test.jsonl \
  --predictions predictions.jsonl \
  --output results/evaluation.json \
  --no-scenario-details

The scorer uses only the Python standard library.

---

Metrics

Final-state accuracy

Item-level accuracy of the predicted final states.

[ \text{FinalStateAccuracy}

\frac{ \text{correct item final states} }{ \text{tracked items} } ]

Final-state exact rate

Percentage of scenarios for which every final item state is correct.

Final-transition accuracy

Item-level accuracy of the predicted principal state transition.

Final-transition exact rate

Percentage of scenarios for which every final transition is correct.

Item-path exact accuracy

Proportion of tracked items whose complete transition path is exactly correct.

Scenario path exact rate

Percentage of scenarios for which every item path is exactly correct.

Step-state accuracy

Accuracy of the predicted item state after each evidence event.

Step-transition accuracy

Accuracy of the transition assigned to each item at each evidence event.

Evidence-role accuracy

Accuracy of item-specific evidence-role assignments.

Decisive-evidence F1

Set-based F1 for identifying decisive evidence events for each item.

Decisive-evidence exact rate

Percentage of scenarios in which all decisive-evidence sets are exact.

Minimal-subset score

A symmetric similarity score comparing predicted and gold minimal decisive subsets.

Exact and partially overlapping evidence subsets receive different scores.

Minimal-subset exact rate

Percentage of scenarios in which every minimal decisive-subset collection is exactly reconstructed.

Changed-item F1

Set-based F1 for identifying items that changed from their prior state.

Preserved-item F1

Set-based F1 for identifying items correctly preserved.

Dependency accuracy

Proportion of gold cross-item dependencies completed at the required evidence point.

Workflow-constraint accuracy

Proportion of workflow constraints satisfied by the predicted state and trajectory.

Complete reconstruction rate

A scenario counts as a complete reconstruction only when all of the following are exact:

• final states;
• final transitions;
• complete paths;
• evidence roles;
• decisive evidence;
• minimal decisive subsets;
• changed items;
• preserved items;
• cross-item dependencies;
• workflow constraints.

This is intentionally strict.

CMESI composite score

The official scorer reports an unweighted mean of:

• final-state accuracy;
• final-transition accuracy;
• item-path exact accuracy;
• step-state accuracy;
• step-transition accuracy;
• evidence-role accuracy;
• decisive-evidence F1;
• minimal-subset score;
• dependency accuracy;
• workflow-constraint accuracy.

The composite is a summary statistic. Individual metrics and complete reconstruction should always be reported alongside it.

---

Matched-pair metrics

The scorer reports:

• paired final-state exact rate;
• paired path exact rate;
• paired complete reconstruction rate;
• mean pair composite score.

A pair is counted as exact only when both its A and B scenarios meet the relevant exactness condition.

Matched-pair performance is important because the two cases often differ by only a small number of decisive details.

---

Diagnostic slices

The scorer aggregates results across named reasoning structures, including:

• evidence-quality override;
• contradiction handling;
• non-monotonic trajectories;
• reconfirmation;
• investigation workflow;
• treatment workflow;
• cross-item cascades;
• historical reinterpretation;
• calibrated ambiguity.

Results are also reported by:

• clinical domain;
• mechanism complexity;
• annotation confidence;
• integration type;
• reasoning-pattern label.

---

Recommended evaluation protocol

For a comparable benchmark result:

1. Use the complete data/test.jsonl scenario inputs.
2. Do not expose gold fields to the evaluated model.
3. Preserve the original evidence order.
4. Require structured JSON output.
5. Do not manually repair predictions before scoring.
6. Record the model name and version.
7. Record the prompt or inference protocol.
8. Record whether examples from the training split were included in context.
9. Run the official scorer without modification.
10. Report the complete metric set, not only the composite score.
11. Report matched-pair performance.
12. Report parsing failures and missing predictions.

---

Creating model inputs without gold annotations

For an evaluation interface, retain:

scenario_id
pair_id
domain
prior_context
tracked_items
prior_state
evidence_sequence

Remove:

gold_final_state
gold_transition_path
gold_final_transition
evidence_role_by_event
evidence_quality_by_event
item_evidence_trace
decisive_evidence
minimal_decisive_subsets
changed_items
preserved_items
cross_item_dependencies
workflow_constraints
valid_alternative_paths
integration_types
reasoning_patterns
integration_structure
mechanism_complexity
complexity_components
annotation_confidence
confidence_rationale

Evidence-quality annotations should be removed for the standard benchmark task because the model is expected to infer quality from the event descriptions.

They may be retained in a separate oracle-quality experimental condition.

---

Suggested prompt

You are given a synthetic clinical state-integration scenario.

Track every listed item across the evidence sequence.

For each evidence event:

1. determine whether the evidence changes each item;
2. assign the resulting state;
3. assign the transition;
4. assign the evidence role for that item.

Then return:

- the final state of every tracked item;
- the complete transition path;
- the principal final transition for each item;
- evidence roles by event;
- decisive evidence for each item;
- minimal decisive evidence subsets;
- changed items;
- preserved items.

Do not collapse the case into a single diagnosis.
Do not omit unchanged items.
Do not treat technically limited or invalid evidence as definitive.
Return valid JSON only.

---

Baseline systems

Useful baseline families include:

Final-state-only baseline

Predicts only final_state.

This baseline cannot receive credit for transition reconstruction, evidence roles, dependencies or workflow compliance.

Last-evidence baseline

Updates item states using only the final evidence event.

This tests whether full-sequence integration adds value.

Monotonic baseline

Allows hypotheses to become more certain or less certain but prevents reactivation and reconfirmation.

This is expected to fail on non-monotonic cases.

Independent-item baseline

Predicts each item independently without cross-item dependencies.

This tests the importance of integrated state architecture.

Equal-evidence baseline

Treats all evidence events as equally adequate and reliable.

This tests evidence-quality sensitivity.

Large-language-model structured baseline

Prompts a general language model to return the complete prediction schema without additional training.

---

Intended uses

CMESI is intended for:

• research on structured clinical reasoning;
• evaluation of multi-evidence integration;
• analysis of state-tracking failures;
• testing non-monotonic reasoning;
• evaluating sensitivity to evidence quality;
• studying workflow and dependency reasoning;
• comparing single-answer and multi-state architectures;
• developing structured-output evaluation methods;
• and investigating matched-pair robustness.

---

Out-of-scope uses

CMESI is not intended for:

• clinical diagnosis;
• treatment selection for real patients;
• medical triage;
• autonomous clinical decision-making;
• measuring the safety of deployment in healthcare;
• estimating performance on electronic health records;
• replacing clinician review;
• medical education without independent verification;
• or making claims of clinical competence from benchmark performance alone.

A high CMESI score does not establish that a system is medically safe, clinically validated or suitable for patient care.

---

Synthetic-data statement

All scenarios were constructed for benchmark purposes.

They do not describe real patients.

The dataset contains:

• no names;
• no addresses;
• no dates of birth;
• no medical record numbers;
• no real clinical notes;
• no protected health information;
• and no linkage to external patient datasets.

The cases use clinically recognisable reasoning structures to create controlled state-transition problems. They should not be interpreted as clinical protocols.

---

Limitations

Synthetic setting

The scenarios are more controlled and explicit than real clinical documentation.

Real records contain:

• missing events;
• uncertain chronology;
• duplicated information;
• inconsistent terminology;
• documentation errors;
• fragmented care pathways;
• and ambiguous responsibility for decisions.

CMESI does not reproduce the full complexity of real records.

Limited scale

CMESI v0.2 contains 48 scenarios.

Its purpose is mechanism-rich evaluation rather than population-scale estimation.

Small differences between systems should not be treated as statistically definitive without additional evaluation.

Constructed ontology

The state and transition ontologies are benchmark abstractions.

Clinical concepts can have more nuanced or institution-specific states than those represented here.

Gold-path specificity

Although the benchmark permits explicit alternatives in selected ambiguous cases, most scenarios use one canonical transition path.

Other defensible narrative explanations may exist even when they do not match the structured gold path.

Domain coverage

The benchmark spans several domains but does not provide comprehensive coverage of any speciality, disease group or demographic population.

No clinical validation

The dataset has not been designed or validated as a clinical decision-support system.

Potential language-model familiarity

Clinical patterns used in the synthetic cases may resemble concepts present in model training corpora. The matched-pair and full-path design reduces but does not eliminate reliance on learned clinical associations.

Annotation dependence

Mechanism labels, complexity values, decisive-evidence sets and minimal subsets are part of the benchmark design. Some boundaries remain judgement-dependent.

---

Bias considerations

Because the benchmark is synthetic and mechanism-focused, it does not systematically represent:

• age;
• sex;
• gender;
• ethnicity;
• disability;
• socioeconomic status;
• healthcare access;
• geographic variation;
• or differences between health systems.

This reduces direct demographic inference but also prevents CMESI from evaluating fairness across patient groups.

Performance should not be generalised to underrepresented or unrepresented populations.

---

Reporting results

A benchmark report should include:

Model:
Model version:
Evaluation date:
Prompting method:
Number of in-context examples:
Temperature:
Maximum output tokens:
Structured-output enforcement:
Prediction parsing method:
Missing predictions:
CMESI composite score:
Final-state accuracy:
Final-state exact rate:
Final-transition accuracy:
Scenario path exact rate:
Step-state accuracy:
Step-transition accuracy:
Evidence-role accuracy:
Decisive-evidence F1:
Minimal-subset score:
Dependency accuracy:
Workflow-constraint accuracy:
Complete reconstruction rate:
Paired final-state exact rate:
Paired path exact rate:
Paired complete reconstruction rate:

Results should also include at least one diagnostic breakdown.

---

Versioning

v0.2

CMESI v0.2 introduces a substantially expanded state-integration architecture, including:

• richer multi-item scenarios;
• larger evidence sequences;
• explicit evidence-quality annotations;
• technical invalidation and repeat-testing structures;
• non-monotonic trajectories;
• reconfirmation;
• cross-item dependencies;
• workflow constraints;
• decisive-evidence annotations;
• minimal decisive subsets;
• valid alternative paths;
• matched-pair scoring;
• mechanism-complexity annotations;
• and an expanded official scorer.

Version 0.2 should be treated as distinct from earlier CMESI releases.

Results from different benchmark versions should not be compared without identifying the version used.

---

Reproducibility

The repository includes:

• complete training data;
• complete test data;
• deterministic gold annotations;
• a standard-library-only scorer;
• validation utilities;
• prediction-template generation;
• per-scenario reporting;
• matched-pair reporting;
• and mechanism-specific diagnostic slices.

The scorer version is included in every generated report.

---

Citation

A formal paper citation is not yet assigned.

Until one is available, cite the dataset as:

@dataset{cmesi_v02,
  title        = {Clinical Multi-Evidence State Integration Benchmark},
  shorttitle   = {CMESI},
  version      = {0.2},
  year         = {2026},
  note         = {Synthetic benchmark for structured clinical state integration},
  license      = {MIT}
}

Replace or extend this entry when the benchmark receives a permanent repository identifier or DOI.

---

License

CMESI v0.2 is released under the MIT License.

See the repository LICENSE file for the complete license text.

---

Medical disclaimer

This dataset is provided solely for research and evaluation.

It does not provide medical advice and must not be used to diagnose, treat or manage any real person.

The clinical content is synthetic and may simplify, combine or abstract real clinical processes for the purpose of evaluating AI reasoning.

Always rely on appropriately qualified healthcare professionals and validated clinical guidance for real medical decisions.

[1]: https://huggingface.co/docs/hub/datasets-cards?utm_source=chatgpt.com "Dataset Cards · Hugging Face"