Open questions
Research
Experiments in progress. Each starts with a question and states, honestly, whether it is open, running, or concluded.
- runningAdversarial verification of agent findingsAn audit agent will confidently report issues that are not real. This is a running experiment in making a second, adversarial pass cheap enough to always run.
- openEvaluating a RAG assistant for grounding, not fluencyA retrieval assistant that sounds confident is easy. One that refuses to answer beyond its sources is the hard, useful version. This is an open experiment in measuring that.