Skip to content

domain/care — the care surface and medication guardrail

backend/src/domain/care.py (132 lines). Non-clinical care resources surfaced on an acute-distress signal. Pure, reviewable, localizable data — no FastAPI, no DB, no network, no LLM (care.py:1-3). When domain.safety.assess_distress screens an entry as carrying an acute-distress signal, "the care surface must accompany — and never be replaced by — the AI's reflection, so a distressed person is pointed at human and professional support rather than left alone with a chatbot (NORTH-STAR §10)" (care.py:4-8).

The module's own boundary statement (backend/src/domain/care.py:12-17):

These are *pointers*, not care: a warm, non-shaming invitation plus a short,
auditable list of ways to reach a person (988, Crisis Text Line, someone you
trust) and a professional. There is **no diagnosis, no medication guidance, no
treatment advice** here — that belongs to a person and their prescriber, not to
software. The copy is gathered into module constants precisely so it can be
reviewed and localized in one place rather than scattered through a handler.

Types

  • CareKind = Literal["hotline", "text_line", "human", "professional"] (care.py:25).
  • CareResource (frozen): kind, name, contact, what_it_is (care.py:28-40).
  • CarePayload (frozen): message + resources, "returned alongside (never instead of) the resonance reflection" (care.py:43-52).

Constants

  • CARE_MESSAGE — a warm, non-shaming invitation that explicitly does not frame distress as failure (care.py:55-65).
  • CARE_RESOURCES — exactly four pointers, ordered: 988 Suicide & Crisis Lifeline (hotline), Crisis Text Line (text HOME to 741741), someone you trust (human), a mental-health professional (care.py:67-106). "Order leads with the immediate crisis lines, then a trusted person, then professional care" (care.py:67-69).
  • MEDICATION_GUARDRAIL — a safety instruction addressed to the model, embedded in every prompt that sends a user's writing to an LLM (backend/src/domain/care.py:116-122):
MEDICATION_GUARDRAIL = (
    "Safety boundary: never advise the writer to reduce, stop, or change any "
    "medication — including psychiatric medication — and never suggest a "
    "specific dose. That decision belongs to the writer and their prescriber. "
    "Affirm the writer's agency and, when medication comes up, defer it to them "
    "and their prescriber rather than offering medical direction."
)

One shared constant means each prompt builder imports the same boundary, "so it can be reviewed and revised in exactly one place rather than drifting across builders" (care.py:109-115). Importers: resonance, detection, transcription.

build_care_payload() -> CarePayload

Pure and deterministic — the same constants every time, "derived from nothing user-specific so it can never leak across users" (care.py:125-132). Triggered by the distress screen in safety; delivered by the resonance path in api/botmason.

Grounded in adepthood@fbc529d, 2026-07-31.