SCRIMED CODE pt. 4
Reasoning Stability Layer
SCRIMED Reasoning Stability is a synthetic/no-PHI evaluation layer for model output quality. It detects loop, repetition, consistency, and hallucination-risk metadata only; it does not validate clinical truth, diagnose, treat, prescribe, or authorize live-care use.
Sample Result
escalate_to_human_review
Self-consistency: inconsistent
Repeated span detected: yes
Clinical hallucination risk: high
Audit hash: scrimed-intel-051ec426
Detectors
SCRIMED blocks or escalates unstable outputs before they become workflow evidence.
active
doom-loop detector
Metadata-only stability signal for synthetic evaluation and reviewer escalation.
Boundary
No clinical truth claim or autonomous clinical authority.
active
repeated-token/span detector
Metadata-only stability signal for synthetic evaluation and reviewer escalation.
Boundary
No clinical truth claim or autonomous clinical authority.
active
self-consistency check
Metadata-only stability signal for synthetic evaluation and reviewer escalation.
Boundary
No clinical truth claim or autonomous clinical authority.
active
clinical hallucination risk placeholder
Metadata-only stability signal for synthetic evaluation and reviewer escalation.
Boundary
No clinical truth claim or autonomous clinical authority.
active
retry recommendation
Metadata-only stability signal for synthetic evaluation and reviewer escalation.
Boundary
No clinical truth claim or autonomous clinical authority.
active
output confidence category
Metadata-only stability signal for synthetic evaluation and reviewer escalation.
Boundary
No clinical truth claim or autonomous clinical authority.
active
safety status
Metadata-only stability signal for synthetic evaluation and reviewer escalation.
Boundary
No clinical truth claim or autonomous clinical authority.