Bom dia, Paul! Sou advogado no Brasil e estou aprendendo muito com seus textos. Infelizmente a diferença de moeda é muito grande, mas futuramente posso virar um assinante ativo para aprender com mais profundidade! Sucesso! @ramizlazarine
This is useful for clinical AI because progressive disclosure is not just token hygiene; it is an authority boundary.
For an OR-adjacent agent, the chart, imaging, preference card, huddle update, and room state should not all be loaded as equal context. The system should fetch the known source deterministically, keep durable rules in explicit artifacts, and route the unresolved conflict to a named human instead of spending reasoning budget turning it into confidence.
Progressive disclosure first is the right order. The one I would add: measure cache reads, not just prompt size. I had a fan-out where every worker dragged a large context through every turn, and the re-reads, not the new tokens, were most of the cost. Nothing looked wasteful line by line. The waste was in how often the same context got read again.
Great tips Paul, thanks for sharing :-)
Thanks, Pau 🤘
Bom dia, Paul! Sou advogado no Brasil e estou aprendendo muito com seus textos. Infelizmente a diferença de moeda é muito grande, mas futuramente posso virar um assinante ativo para aprender com mais profundidade! Sucesso! @ramizlazarine
This is useful for clinical AI because progressive disclosure is not just token hygiene; it is an authority boundary.
For an OR-adjacent agent, the chart, imaging, preference card, huddle update, and room state should not all be loaded as equal context. The system should fetch the known source deterministically, keep durable rules in explicit artifacts, and route the unresolved conflict to a named human instead of spending reasoning budget turning it into confidence.
Progressive disclosure first is the right order. The one I would add: measure cache reads, not just prompt size. I had a fan-out where every worker dragged a large context through every turn, and the re-reads, not the new tokens, were most of the cost. Nothing looked wasteful line by line. The waste was in how often the same context got read again.