BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//aiweek.boston//Schedule//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:The Model Picks the Code\, the Rules Decide: Guardrails for an LLM-Driven Medical Interview
X-WR-TIMEZONE:America/New_York
BEGIN:VTIMEZONE
TZID:America/New_York
X-LIC-LOCATION:America/New_York
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:20070311T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:20071104T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:the-model-picks-the-code-the-rules-decide-guardrails-for-an-llm-driven-medical-i@aiweek.boston
DTSTAMP:20260918T222410Z
SEQUENCE:0
STATUS:CONFIRMED
DTSTART;TZID=America/New_York:20260930T090000
DTEND;TZID=America/New_York:20260930T100000
SUMMARY:The Model Picks the Code\, the Rules Decide: Guardrails for an LL
 M-Driven Medical Interview
DESCRIPTION:MedBot is a research prototype that conducts structured medic
 al interviews with synthetic patients in simulated cases\, then hands the
  resulting report to a physician evaluator. It never gives advice to pati
 ents.\n\nWhat you'll see:\n\n- Closed catalogs instead of free-text safet
 y decisions. The model can select only red-flag codes from a YAML catalog
 . Urgency levels and escalation language come from deterministic rule fil
 es\; out-of-catalog values are discarded as invalid.\n- Structured output
  generated from the rules. JSON Schema enum values are generated from the
  same rule files used during runtime validation.\n- Detect\, then audit. 
 One LLM call selects red-flag codes\; a second call audits each code agai
 nst its formal condition. Giving the condition to the detector itself dro
 pped sensitivity from 1.00 to 0.91\, so the two jobs stay separate. If ei
 ther call returns no verdict\, the turn is rolled back.\n- Measure\, don'
 t guess. A phrase matcher achieved 0.81 sensitivity on the tuning set but
  only 0.60 on a sealed holdout set. Across a 220-case evaluation\, an ope
 n-weight model detected every labeled alarm for approximately $0.05 per r
 un\, although it produced more false positives. The tested Sonnet configu
 ration missed five alarms.\n- Safety gates we removed. A dosage filter co
 uld not reliably distinguish medication history ("the patient takes aspir
 in") from treatment advice ("take aspirin"). A fact-grounding gate reject
 ed 25 of 40 first-pass outputs.\n- Prompt caching in a multi-turn intervi
 ew. Why the framework's default caching strategy yields zero cache reads\
 , and costs more than running without caching.\n\nWho it's for: engineers
  building LLM features in regulated or high-stakes domains\, including he
 althcare\, finance\, and legal services. The implementation uses Java and
  Spring AI with Anthropic and open-weight models\; the patterns are stack
 -agnostic.\n\nIn person\, 16 seats\, 11th floor (Singapore room). Doors 8
 :45 AM\; we start at 9:00 sharp and end at 10:00 AM.\nRegister on Luma: h
 ttps://luma.com/h3mw9ge8\n\nHosted by Piotr Obolewicz · https://obolewicz
 .com
LOCATION:- CIC Cambridge\, One Broadway\, 11th floor (Singapore room)\n- 
 Doors 8:45 AM\; we start at 9:00 sharp and end at 10:00 AM\n- 16 seats. I
 f your plans change\, please cancel your RSVP so someone on the waitlist 
 can take your seat.\, 1 Broadway\, Cambridge\, MA 02142
URL:https://aiweek.boston/schedule/the-model-picks-the-code-the-rules-dec
 ide-guardrails-for-an-llm-driven-medical-i
BEGIN:VALARM
ACTION:DISPLAY
DESCRIPTION:The Model Picks the Code\, the Rules Decide: Guardrails for a
 n LLM-Driven Medical Interview
TRIGGER:-P1D
END:VALARM
BEGIN:VALARM
ACTION:DISPLAY
DESCRIPTION:The Model Picks the Code\, the Rules Decide: Guardrails for a
 n LLM-Driven Medical Interview
TRIGGER:-PT1H
END:VALARM
END:VEVENT
END:VCALENDAR