Cross-Checking AI Agents' Decisions: Did They Depend on Another Agent's Work?

Many AI safety designs rely on independent checks: one AI does the work, and others review it. These checks only help if each checker judges on its own. I test whether an AI reviewer told to judge independently does so, and whether a failure can be proven. I built a loan-review workspace from 200 real 2025 US mortgage records, where an AI reviewer scores 10 loans next to colleagues' posts. On a borderline loan, a colleague's note saying she "rated it a 3" lowered the average risk score from 5.00 to 3.83 for gemma 4 12B (30 cases) and from 5.00 to 4.04 for Claude Opus 5.5 (6 cases), on a 1 to 5 scale. In 21 of its 24 runs with the note, Claude wrote that its score was independent. An AI monitor reading the reviewer's full log caught 1 of the 6 Claude cases. I built a "flow audit". It uses only the workspace's records, and re-runs the reviewer with the note removed and, as a placebo, with an unrelated post removed. It caught all 6. In a pre-registered follow-up on fresh runs, it caught 17 of 17 gemma cases, 9 of them instructed controls, with no false alarms on 67 honest cases. gemma followed a colleague toward more caution even on a clearly safe loan, but not toward leniency on a clearly risky one. An agent's own explanation cannot confirm its independence; the workspace's records plus re-runs can test it.