domain/safety — acute-distress screening

backend/src/domain/safety.py (375 lines). A pure, deterministic screen — "not a diagnosis, not treatment, not advice" (safety.py:1-10). assess_distress(text) returns a typed DistressSignal(level, category) where level is "elevated" only for explicit acute phrasing in four categories (DistressCategory, safety.py:83-90): suicidal_intent, self_harm, medication_cessation, intent_to_harm. Consumed by the journal care path: POST /journal/{entry_id}/resonance screens the entry and, on an elevated signal, returns the care surface alongside the reflection (safety.py:8-10).

Deliberately conservative

The design statement (backend/src/domain/safety.py:14-23):

Adepthood honors ordinary darkness — grief, sadness, emptiness, the "dark night
of the soul", existential struggle — as real developmental territory, and must
never pathologize it (NORTH-STAR §10). So matching is intentionally narrow: each
category fires only on phrasing that expresses *acute intent or action* ...
When phrasing is ambiguous between ordinary darkness and acute distress we
prefer **not** to flag ... The phrase lists below are explicit and auditable
rather than clever; add to them only with a corresponding negative test
guarding ordinary darkness.

The phrase lists (safety.py:113-158)

Explicit regex alternations per category, matched case-insensitively on word boundaries — e.g. kill myself, end my life, want to die, better off dead (suicidal intent); hurt myself, self[ -]harm (self-harm); kill (him|her|them|someone|everyone) (intent to harm); stop(ping|ped)? (taking )?(my )?(meds|medication|pills|antidepressants) (medication cessation). Categories are checked in list order; the first non-denied match wins (safety.py:108-112,371-374).

Normalization (_normalize, safety.py:328-343)

Pipeline order is load-bearing: (1) fold six alternate apostrophe code points to ASCII ' — before NFKC, which would decompose U+00B4 into space + combining acute (safety.py:165-177); (2) neutralize every Unicode format character (category Cf: zero-width, bidi controls, soft hyphen) to a space so "invisible splices cannot defeat matching" (safety.py:179-181); (3) NFKC (folds fullwidth forms); (4) lowercase + whitespace collapse. Rationale: "the dangerous failure direction for a crisis screen is a false negative" (safety.py:29-32).

Negation handling (deterministic, no parser)

• Window: a match is suppressed only by a negator in the last _NEGATION_WINDOW_WORDS = 5 tokens strictly before it, clipped to the current clause — boundaries are . ! ? ; , and the coordinating conjunctions and/but (safety.py:183-218,231-241). or/nor deliberately do not clip: "I would never kill myself or hurt myself" is a single denial (safety.py:203-209).
• Logical negators (never, not, no plans to, don't, …) count by parity: odd negates, even (including zero) does not — double negation ("I can't say I don't want to die") still flags (safety.py:186-195,244-256). Bare no is excluded: it would negate "No. I want to kill myself" (safety.py:188-189).
• Temporal negators (no longer, used to) negate only when directly abutting the match — "more than I used to I want to die" still flags (safety.py:197-201).
• Or-chain inheritance: a match that is the bare right conjunct of an "X or Y" / "X nor Y" coordination is suppressed when its left acute conjunct is itself denied, resolved iteratively with memoization so "an adversarially long or-chain terminates without RecursionError" (safety.py:58-67,259-325); each walk step moves strictly leftward, and caching collapses overlapping walks "from cubic to quadratic total work" (safety.py:304-308). Any intervening words break the coordination — "I have no plans to kill myself or I will hurt myself" still flags the self-harm conjunct (safety.py:65-67).
• A negator that is part of a positive phrase (the "don't" in "don't want to be alive anymore") can never self-suppress, because only text strictly before the match is inspected (safety.py:54-56,247-248).

Worked examples

Input	Result
"I feel empty and lost in the dark night of the soul"	none — ordinary darkness passes through
"I want to die"	elevated / suicidal_intent
"I would never kill myself"	none — one logical negator in the window
"I can't say I don't want to die"	elevated — even count is not, odd is; here two negators = even… the parity rule counts can't + don't = 2 → not negated → flags
"I would never kill myself or hurt myself"	none — the denial is inherited across or
"I would never hurt myself, but I want to kill him"	elevated / intent_to_harm — the comma and but clip the window
"I used to cut myself"	none — abutting temporal negator

Purity: standard library only — no FastAPI, SQLModel, network, or LLM imports (safety.py:69).

---

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