Konsil: An Adversarial AI Medical Case-Review Board

Описание и применение

Konsil – это открытая система многопользовательского обзора медицинских случаев, где агенты сохраняют разногласия, проверяют доказательства и проходят аудит, предотвращая ошибки и гипотезы.

Задача

Multi‑agent systems often converge too quickly, reinforcing shared assumptions and producing overconfident answers.

Результат

Konsil’s architecture prevented hallucinated author names, silent removal of disagreements, and composite quotations during blind development runs, demonstrating that a system can self‑audit its correction process.

Рабочий процесс

• 1. Blind specialist agents independently review the case.
• 2. Challenger agent questions assumptions and seeks alternative explanations.
• 3. Checklist Officer verifies that all findings are addressed and evidence is cited.
• 4. Steward evaluates cost‑aware sequencing and value of additional information.
• 5. Dissent Ledger records all disagreements with reasoning and confidence.
• 6. Reviewer Gate audits citation integrity, dissent integrity, safety language, and completeness; failures trigger corrections and re‑audit.

Доказательства

Agents can converge too quickly, reinforce shared assumptions, overlook alternative explanations, and produce an answer that sounds more confident than the available evidence justifies.

Источник

After the initial assessments, a dedicated Challenger agent reviews the emerging positions. Its job is to question assumptions, explore alternative explanations, identify missing evidence, and challenge premature conclusions.

Источник

The Reviewer Gate caught: Hallucinated author names in otherwise valid references. A disagreement that had been silently removed from the final minutes. A composite quotation incorrectly presented as verbatim.

Источник