Discussion about this post

User's avatar
Augmenting_decision's avatar

Self-evaluating reports make me nervous as a trust mechanism, the agent grading its own homework can drift into confident-sounding nonsense. What's the actual check that catches a plausible but wrong report before it ships?

Abhishek Pan's avatar

A single new passage_id keeps the loop going, even if it landed at the bottom of the RRF list. Counting only new passages above a score floor stops the runs that keep paying for a reflection step to add one weak hit per round.

Ready for more?