Twelve benchmarks. One shared question: Has the AI exceeded what it can justify?
Domain: Ambulatory triage and incomplete clinical evidence
Core question: What conclusions and actions are justified before the diagnosis is complete?
Tests whether a model can recognize diagnostic uncertainty while still selecting a safe, evidence-proportional triage action.
Evidence pressure: Incomplete histories; provisional queue labels; pending or absent evidence; later notes that may be authoritative, partial, irrelevant, or merely operational.
Decision classes: Emergency evaluation; urgent/same-day assessment; routine care; home monitoring; targeted verification; reassessment.
Distinctive contribution: Object-specific confidence and high action confidence despite low diagnostic confidence.
Domain: Early infection, sepsis physiology, and pending diagnostic data
Core question: What stabilization, monitoring, escalation, or de-escalation is justified while source and organism remain uncertain?
Separates confidence in infection, sepsis physiology, source, risk, and immediate action while evidence evolves.
Evidence pressure: Pending cultures and ANC; nonspecific lactate or pyuria; preliminary cultures; competing explanations; later deterioration or authoritative reversal.
Decision classes: Stabilize; urgent assessment; monitored care; de-escalate; transfer; continue evaluation; reassess.
Distinctive contribution: Decision-blocking uncertainty, dynamic reassessment, and confidence revision as physiology changes.
Domain: Medication history, dosing, administration, and discharge reconciliation
Core question: Which medication objects are sufficiently known to support active status, dose, schedule, administration, or discharge instructions?
Requires separate judgments about whether a medication is active, correctly dosed, safely timed, and appropriate to give now.
Evidence pressure: Imported lists; pharmacy fills; patient recall; specialist plans; duplicate entries; completed courses; high-risk dosing; partial later clarification.
Decision classes: Continue; hold; correct; clarify; administer; defer; discontinue; escalate; reassess.
Distinctive contribution: Granular justification boundaries across medication status, dose, schedule, risk, and action.
Domain: Multidisciplinary specialist disagreement and plan reconciliation
Core question: Is the conflict real, already reconcilable, or compatible with a safe interim plan?
Tests whether a model can interpret recommendations within each specialist's scope, risk object, and time horizon rather than choosing by authority alone.
Evidence pressure: Different clinical objectives; conditional recommendations; stale notes; partial agreement; later operational updates; unresolved risk tradeoffs.
Decision classes: Reconcile; select a bounded plan; define interim action; request joint review; defer one component; monitor; escalate.
Distinctive contribution: Qualified disagreement, domain-specific confidence, and reconciliation without false consensus.
Domain: Interfacility transfer and incomplete external records
Core question: Which missing information truly blocks action, and which gaps should not delay stabilization or capability-based transfer?
Evaluates safe action under incomplete transfer documentation while distinguishing critical unknowns from non-blocking record gaps.
Evidence pressure: Missing medication history; absent imaging; incomplete handoff; unverifiable procedures; outside reports; patient instability; destination capability constraints.
Decision classes: Stabilize; accept; transfer; verify; hold a specific action; continue monitored care; defer final decision; reassess.
Distinctive contribution: Action-specific documentation thresholds and decision-blocking versus non-blocking missing information.
Domain: Diagnostic test provenance, validity, and evidentiary quality
Core question: Is the available diagnostic evidence reliable enough to support the diagnosis, causal attribution, or treatment?
Tests whether evidence quality is evaluated separately from clinical risk and whether weak or mismatched results are confirmed before overreach.
Evidence pressure: Specimen quality; bedside versus central laboratory results; wrong-patient or mismatched sources; preliminary reads; low-specificity tests; treatment response.
Decision classes: Accept; repeat; confirm; treat provisionally; withhold diagnosis-specific treatment; de-escalate; escalate; reassess.
Distinctive contribution: Confidence in evidence quality separated from confidence in diagnosis, risk, and immediate safety action.
Domain: Patient identity, record matching, and action-specific verification
Core question: Is identity certainty sufficient for the specific action being considered?
Applies different evidence thresholds to chart location, record merge, medication, transfusion, surgery, oncology treatment, and other high-risk actions.
Evidence pressure: Duplicate records; demographic overlap; conflicting identifiers; accession mismatch; aliases; merged histories; partial matching; urgent care pressure.
Decision classes: Locate; verify; merge; hold; treat; transfuse; operate; escalate identity review; use temporary safeguards.
Distinctive contribution: Action-specific identity thresholds and prevention of confidence transfer from low-risk matching to high-risk treatment.
Domain: Pathology interpretation and treatment readiness
Core question: What diagnosis, subtype, staging, and treatment-readiness claims are justified before pathology is fully final?
Separates confidence in malignancy, subtype, staging, treatment urgency, and readiness for specific therapy.
Evidence pressure: Preliminary pathology; pending stains or molecular studies; discordant specimens; urgent complications; later addenda; benign final results; incomplete staging.
Decision classes: Proceed with supportive or staging work; await final pathology; start time-sensitive treatment; hold subtype-specific therapy; revise; reassess.
Distinctive contribution: Partial readiness: some actions may proceed while diagnosis-specific treatment remains unjustified.
Domain: Late-arriving laboratory, imaging, culture, and specialist evidence
Core question: How much should the conclusion, confidence, and action change after new evidence arrives?
Tests proportional updating rather than anchoring, hindsight bias, reflex reversal, or failure to revise.
Evidence pressure: Late positive and negative results; contamination; irrelevant abnormalities; final reads; treatment response; evidence that changes one object but not another.
Decision classes: Maintain; increase; decrease; narrow; reverse; continue action; stop action; schedule or trigger reassessment.
Distinctive contribution: Proportional confidence revision and resistance to both anchoring and overreaction.
Domain: Escalation, destination selection, and level-of-care decisions
Core question: Does risk justify escalation despite diagnostic uncertainty, and what level or destination is proportionate?
Requires models to separate confidence in diagnosis from confidence in immediate risk and the need for higher-level care.
Evidence pressure: Vascular, neurologic, obstetric, oncologic, cardiopulmonary, bleeding, and infectious risk with incomplete diagnostic closure.
Decision classes: No escalation; urgent evaluation; monitored bed; step-down; ICU; operating room; specialty or tertiary transfer.
Distinctive contribution: High confidence in escalation can coexist with low confidence in the exact diagnosis.
Domain: Serious illness prognosis, treatment benefit, function, survival, and goals of care
Core question: What can actually be justified about prognosis, and what should remain uncertain?
Tests whether models avoid overstating survival, recovery, treatment benefit, futility, or timing while still supporting useful goals-of-care decisions.
Evidence pressure: Population statistics versus individual prognosis; uncertain recovery; advanced-therapy candidacy; time-limited trials; DNR misinterpretation; hospice overreach.
Decision classes: Communicate bounded prognosis; continue treatment; define a time-limited trial; involve palliative care; discuss hospice when appropriate; reassess.
Distinctive contribution: Prognostic object and time-horizon calibration with useful recommendations despite irreducible uncertainty.
Domain: Discharge, observation, admission, transfer, and disposition deferral
Core question: When is residual uncertainty compatible with a safe disposition?
Serves as the suite capstone by combining incomplete evidence, competing risks, destination capability, object-specific confidence, and reassessment.
Evidence pressure: Medical stability versus destination safety; unresolved mechanism; time-limited observation; capability gaps; late nonspecific results; evolving trajectory.
Decision classes: Discharge; observe; admit; transfer; defer disposition; revise after observation or new evidence.
Distinctive contribution: Capstone disposition judgment: accepting manageable residual uncertainty without under- or over-escalation.