Skip to content

UCM_GLOSSARY.md

Latest commit

 

History

History
166 lines (106 loc) · 4.96 KB

UCM_GLOSSARY.md

File metadata and controls

166 lines (106 loc) · 4.96 KB
Raw

UCM Glossary

Uncertainty & Confidence Management

This file is the canonical vocabulary source for the UCM benchmark suite.

Benchmarks should reference this glossary rather than redefining shared terms. A benchmark should introduce a new term only when the concept is genuinely new. Any new reusable term should be added here before it is treated as suite-wide vocabulary.

Core Suite Language

Decision-Blocking Uncertainty

Uncertainty that prevents the AI from safely selecting or justifying an action until additional evidence becomes available.

Decision-blocking uncertainty changes what may safely happen now.

Examples:

  • the dose of a high-risk medication is unknown
  • patient identity is not verified
  • a pending result determines whether discharge is safe
  • conflicting evidence changes the treatment or care setting

Non-Blocking Uncertainty

Uncertainty that remains unresolved but does not prevent a safe, justified action.

Examples:

  • the exact organism is unknown, but emergency stabilization is already justified
  • the exact viral cause is unknown, but home care is appropriate
  • final pathology is pending, but the immediate action does not depend on it

Conditionally Blocking Uncertainty

Uncertainty that is not decision-blocking now but becomes decision-blocking if a defined trigger, threshold, or new event occurs.

Examples:

  • a pending culture becomes positive
  • symptoms worsen
  • blood pressure falls
  • the patient fails to improve
  • a source is shown to be mismatched

Confidence Object

The specific object to which a confidence judgment applies.

Common UCM confidence objects:

  • evidence quality
  • conclusion or diagnosis
  • causal attribution
  • risk assessment
  • recommended action

Confidence may differ across objects.

Justification Boundary

The strongest conclusion, confidence level, or action the available evidence supports without overreach.

The justification boundary defines where supported reasoning ends and speculation begins.

Scheduled Reassessment

A planned review at a defined time, result, or checkpoint.

Examples:

  • repeat lactate in two hours
  • review culture results when posted
  • repeat blood pressure after rest
  • review a tacrolimus trough before the evening dose
  • specialist follow-up within seven days

A scheduled reassessment should identify:

  • timing or expected result
  • owner
  • confidence object likely to change
  • action threshold

Triggered Reassessment

A review initiated by a new event, symptom, deterioration, or threshold crossing.

Examples:

  • hypotension
  • new confusion
  • recurrent fever
  • increased oxygen requirement
  • new bleeding
  • failure to improve
  • identity mismatch

A triggered reassessment should identify:

  • trigger
  • urgency
  • owner
  • action that changes

Supporting Terms

Reassessment Threshold

The specific evidence change that should alter the conclusion, confidence, or action.

Evidence-Proportional Action

An action whose urgency and scope match both the evidence and the potential harm.

Residual Uncertainty

Uncertainty that remains after a justified conclusion or action has been selected.

Residual uncertainty may be non-blocking, conditionally blocking, or decision-blocking.

Provisional Conclusion

A conclusion supported for the current moment but explicitly subject to revision when expected evidence arrives.

Confidence Transfer Error

A reasoning error in which confidence in one object is incorrectly transferred to another.

Examples:

  • high confidence in action is treated as high confidence in diagnosis
  • high evidence quality is treated as high causal certainty
  • low diagnostic confidence is used to justify weak action despite high risk

Required UCM Reasoning Framework

Evidence → Uncertainty → Justification → Confidence → Action → Reassessment

Recurring UCM Sentence

The safest conclusion is the strongest conclusion the evidence can justify.

Vocabulary Governance

  1. This glossary is the canonical source for suite-wide UCM terminology.
  2. Benchmark prompts, grading guidelines, and Design Notes should reference this file rather than restating shared definitions.
  3. A benchmark may define a domain-specific term locally when that term is not intended to apply across UCM.
  4. A new reusable UCM term should be added here before it appears across multiple benchmarks.
  5. Changes should preserve earlier meanings unless the glossary version explicitly documents a revision.
  6. Grading should use glossary definitions consistently across submissions.

Output Architecture Terms

Cross-Section Consistency Format

Patients or cases recur across separate evidence, uncertainty, conclusion, confidence, action, and reassessment sections.

Use when cross-section consistency is part of the evaluation.

Patient-Centered Synthesis Format

Each patient or case receives one consolidated reasoning block, followed by a final urgent-handoff section.

Use when the benchmark should minimize duplication and output-fatigue confounds.