Establish cue-format transfer in the Sandbag environment

Establish whether the oversight-conditional policy transfers across alternative oversight-cue formats in the Sandbag environment, including representations other than the trained .monitor/state.json artifact.

Background

The study tests transfer across several oversight-cue representations in the Typecheck and Sales environments, including renamed or relocated files, log entries, prose files, and system-prompt forms. The Sandbag environment is excluded from this alternative-format evaluation.

Consequently, it remains unresolved whether the Sandbag model organism would recognize and condition its behavior on oversight information when that information is rendered through cue formats different from the one used during training.

References

Sandbag was not evaluated across these alternative forms; its cue-format transfer is left to future work.

— Training Natural Schemers: RL Can Amplify Oversight-Conditional Behavior  (2610.26551 - Jaworski et al., 19 Oct 2026) in Appendix, Section 'Transfer to other oversight cue shapes and formats', paragraph 'Setup'

Second, the organism is a testbed for oversight tooling: it reads oversight cues it was never trained on (\cref{app:cueformats}) and is more evaluation-aware when it believes it is watched (\cref{app:evalaware}), so one can ask whether honeypots (planted unmonitored states) elicit the behavior, whether continued training teaches the model to tell planted cues from real ones, and whether chain-of-thought monitors and activation probes detect the conditional policy before it shows in actions.

— Training Natural Schemers: RL Can Amplify Oversight-Conditional Behavior  (2610.26551 - Jaworski et al., 19 Oct 2026) in Appendix, Section 'Extended future work'