CARC Platform — Evidence & Runtime Metrics
The frameworks behind
CARC's Evidence Engine
Technical frameworks for monitoring, governing, and reconstructing clinical agent behaviour — implemented by CARC's Evidence Engine and Runtime Metrics module, and documented here independently of the product.
Decision Context Record™
The Decision Context Record™ captures the runtime context required to reconstruct what a clinical agent recommended, why it acted, what the clinician saw, and how the workflow resolved. CARC's Evidence Engine produces one automatically for every agent run.
Executive Summary
The evidence gap in clinical agent behaviour.
Healthcare AI companies can often demonstrate that a model produced an output. Far fewer can reconstruct how a clinical agent behaved inside the workflow where that output mattered.
The Decision Context Record™ addresses this evidence gap by capturing the minimum operational context required to reconstruct agent behaviour, clinician intervention, and final human action after the fact. It is the record CARC's Evidence Engine produces at runtime.
The framework is designed for healthcare AI companies, SaMD providers, product teams, clinical AI platform teams, and regulatory or quality leaders preparing for EU AI Act healthcare AI, MDR AI software, audit, or post-market AI monitoring.
Framework Components
The Six Components of a Decision Context Record™
Each component captures a discrete category of runtime evidence required to reconstruct clinical agent behaviour in full — and each maps directly to a field CARC's Evidence Engine records.
Agent Output Context
Recommendation, action, escalation, deferral, alert, confidence score, probability estimate, model version, inference timestamp.
Clinical Human Context
Clinician identity or role, authorization level, responsibility, review status, and available intervention authority.
Workflow Environment
Patient information, supporting clinical evidence, risk indicators, warnings, alerts, thresholds, and explanatory content displayed.
Intervention Context
Available actions, override options, escalation pathways, deferral rules, and request-for-review mechanisms.
Clinical Outcome Path
Accepted, overridden, modified, escalated, deferred, rejected, or converted into another workflow action.
Interaction Evidence
Timestamps, interaction sequence, response latency, user actions, workflow navigation, and repeated-case consistency signals.
Reconstruction Test
The Decision Reconstruction Test™
CARC's Decision Replay and Runtime Metrics exist to pass a practical test for clinical AI oversight evidence:
“Could an independent reviewer reconstruct what the clinical agent did six months later?”
Recommendation
Reconstruction
Context
Reconstruction
Intervention
Reconstruction
Responsibility
Reconstruction
Oversight
Reconstruction
Maturity Model
Runtime Oversight Maturity Model™
Five levels of operational evidence maturity for clinical AI agents running in regulated healthcare environments — the same levels CARC's Runtime Metrics module measures against.
Output Logging
The system records that an output, alert, recommendation, escalation, or deferral occurred.
Agent Traceability
The company knows what the agent did and where it happened in the clinical workflow.
Runtime Context Capture
The system captures the clinical, model, threshold, workflow, and UI context that shaped behaviour.
Decision Reconstruction
The full path from agent recommendation to final human action can be reconstructed.
Verifiable Runtime Control
The company can independently demonstrate effective oversight, intervention, and consistent behaviour after deployment — the target level CARC is built to support.
Regulatory Relevance
Designed for runtime evidence in regulated healthcare AI.
The framework supports evidence generation for effective human oversight, clinical AI auditability, AI runtime monitoring, decision reconstruction, and post-market AI monitoring in regulated healthcare AI environments.
The Decision Context Record™ is not a complete compliance solution and does not constitute legal or regulatory advice. It addresses a specific evidential gap: the operational proof required to demonstrate what a clinical agent did, why it acted, what context it used, and how human oversight operated at the point of use. Organizations should seek independent legal and regulatory counsel for specific compliance matters.
See It In CARC
These frameworks run
inside CARC
If your AI product acts inside clinical workflows, the critical question is not only whether a human remains in the loop. It is whether your company can demonstrate what the agent did, why it acted, and who intervened. Request a demo to see CARC generate this evidence automatically.