Mental Health

A single line from an intake note can cost someone a job, a custody arrangement or a place to live. Behavioral health teams know this, which is why they hesitate before sending anything to a model. This starter turns that hesitation into a concrete set of checks. It masks identifiers, scores the reply, and captures metadata rather than the words themselves.

Behavioral health records carry a second layer of law. Substance use disorder records in the United States sit under a separate federal confidentiality rule that limits redisclosure. A privacy officer therefore treats one leaked line as a serious event. That is the reason to place controls in the request path rather than in a later review.

Screening for injected instructions opens the chain and keeps a crafted prompt from disarming later checks. Identity enforcement then blocks a caller the gateway cannot verify. Health record heuristics sit above the general masking stage. Masking covers patient numbers, member numbers and insurance or policy numbers. None of it detects a disclosure written in plain narrative, and a clinician can still paste one.

Captured history keeps metadata only. That choice matters here more than in most settings, because the text of a behavioral health prompt is exactly what a team must not accumulate. Blocked requests are still recorded, so a refusal leaves a trace. The capture setting does not fail open.

On the response side the numbers are strict. A reply must clear 0.88 combined, with relevance at 0.78 and accuracy at 0.80, or the gateway refuses it. Short output under 120 characters fails without reaching a score. Source checking marks a weak citation and still delivers the text. The oversight point holds no rule yet, so crisis routing or clinician sign off is something you must configure.