What happens when we change the world around an agent?
Eight real Codex trials worked through captured production responses. The verifier rejected valid quotations. We kept every original verdict and assessed every answer separately with the repaired verifier.
The later assessment is post hoc. Neither set of counts establishes a performance improvement.
Read the experiment →A fetched document is not the interface every agent sees. Before treating a finding as a performance claim, identify the agent, its observations and the task it must complete.
A source index changed how Codex found authentication facts. The success comparison stayed inconclusive, while the retained traces revealed recovery work the final answer concealed.
Reconciling a dispatch means resolving disagreements between files. This synthetic operations task makes those consequences verifiable; its first comparison also shows the limits of a small study.