Before a patient leaves the hospital, 100 adversarial reviewer agents each get one job: find the documented way this discharge plan fails this patient — with verbatim chart receipts — or be refuted in review. A strict four-test judge (evidence, addressed, materiality, actionability; refute by default) rejects ~95% of objections on screen, with reasons: alert fatigue is filtered at the review gate, never forwarded to the physician. The few surviving findings render as evidence cards anchored to the exact chart lines that prove them, each with one staged order. The clinician accepts or dismisses-with-documented-rationale; accepted orders amend the plan; a targeted re-verification challenges only the amended dimensions: "0 surviving objections — attacked and not broken. Not a safety certification." Every receipt is machine-verified verbatim in code, so fabricated evidence cannot render, and every number on screen is measured, never simulated. The same swarm reviews any encounter from the organizer's synthetic-ambient-fhir-25 dataset (or a pasted/edited discharge summary), and every run exports CDS Hooks cards — the exact payload an Epic order-sign hook would render at discharge signing. Haiku 4.5 reviewer swarm + Opus 4.8 judge over a prompt-cached chart; ~$2 and ~2.5 minutes per 100-reviewer run. Built entirely today: pipeline, judge, CLI, and the EMR-style mission log, all rendering one live event stream.
Team
Placement
Participant
