Confidence-to-Care is a logic-grounded framework that converts confidence signals from medical LLMs into auditable care actions. Token-level log-probabilities are summarized with sequence-level statistics and augmented by robustness cues-semantic entropy, self-consistency across sampled reasoning paths, and agreement across paraphrases. A transparent Prolog policy maps to Care-0-Care-3 via thresholds and duty-to-warn predicates, emits a human-readable trace, and aligns outputs with ESI/CTAS/NEWS2 for recognizable escalation levels. Evaluations on a clinical QA/triage bank and synthetic red-flag cases indicate improved calibration, reliable selective prediction, and appropriate escalation of high-risk cases with negligible overhead. By explicitly linking confidence to care through executable rules, Confidence-to-CCare offers a practical path to trustworthy, auditable, and cost-aware clinical AI for human-in-the-loop workflows. Artifact & Code (v0.1.0): https://github.com/ShabnamAtf/confidence-to-care/releases/tag/v0.1.0