- Arabic masculine/colloquial emergency phrases escalate (مش قادر اتنفس, بختنق);
new avulsed-tooth, swallowed-object, immunocompromised, fever+swelling rules;
idiom/third-party false positives gated
- Agent panel reports real per-run telemetry (Complete/Partial/Fallback, live
flag count); cached examples recompute red flags through the live rules engine
- First paint auto-loads the de-identified real case; mobile columns stack with
44px touch targets; bilingual handoff renders side-by-side; PDF matches the
app's brand tokens and preserves punctuation
- Eval: 60/60 recall, 17/17 specificity, 1,540/1,540 mutation audit, 168 tests
- Internal working documents removed from the public Space
- All AI calls (interview, extraction, handoff enrichment) now run in-Space
via @spaces.GPU — no external endpoint, no secrets required
- Patient name flows through intake, examples, and the handoff card
- Deterministic rules gain Egyptian Arabic colloquial triggers, acutely
qualified so routine history never trips an emergency rule
- Judge-reproducible evals shipped in scripts/: 50/50 recall, 11/11 benign,
1,220/1,220 mass audit; 128 local tests
Close safety-grill false negatives and harden model-claim guard
Lay/contraction phrasing now fires the airway, neurologic, abscess, trauma,
bleeding, pain, crown, and GCA rules; denial contractions added to negation
(can't/won't deliberately excluded — inability phrasing is the symptom);
hedged-diagnosis and soft-directive patterns plus Unicode normalization in
the model-claim guard; cache-load guard parity; age input clamped; prompt
trimmed to writable keys. Tests 41 -> 71. Audit record in
GRILL_ROUND2_RESULTS.md.
