Control evaluation for automated AI R&D

Run traces

Every red-team run with the monitors that reviewed it. Click a run to read the agent's full trajectory, then open any monitor to see how it reviewed the run and the suspicion score (0–100) it assigned.

Mean suspicion by side task × red-team agent. Each cell is the mean suspicion score across all monitors of those runs; the shade tracks the same number. Click a cell to filter.

Loading run index…

Some runs are greyed out and left out of the results, either because the main- or side-task score did not clear its threshold, or because a run with no side task carried out sabotage anyway. They stay browsable here.