Research agents log what they cite — should they also defend what they reject?

Entire_Mark8010 · reddit · 2026-09-15

The author argues that auditable research agents should record not just citations but rejected sources and exclusion reasons. A worked example: an agent reviewing clinical studies silently dropped two null-result studies due to a table-parsing failure, making a pipeline bug look like deliberate curation. Like GKE's GPU dutycycle metric, a citation list shows what entered the answer but not what happened to the rest — an agent rejecting 30 papers can produce the same reference list as one rejecting 3. The author proposes logging exclusions (incompatible metrics, missing columns, extraction failures, unresolved conflicts) and asks how evaluators should split budgets between validating citations and sampling rejections. Discloses a partnership with Apodex, whose Task Board visualizes task breakdown and subagent activity.

Original post →

More from coding & agent

coding & agent channel →