The "type absence — ok/absent/failed instead of a plausible null" decision is the right call and under-appreciated. I run a large multi-agent setup and the fail

The "type absence — ok/absent/failed instead of a plausible null" decision is the right call and under-appreciated. I run a large multi-agent setup and the failures that cost me most weren't wrong answers, they were confident answers over gaps the model should have refused. Curious how you handle it when two sources disagree rather than one being absent — do you surface both with provenance and let the agent decide, or pick a winner server-side?