Evaluation of prompt-free detector selection

Evaluate rule recovery, detector-equivalent configuration selection, and detection power when the watermark detector must be selected from the generated response without access to the original prompt or trusted generation metadata.

Background

The framework’s experiments use known-configuration detection: each response is scored with the detector associated with the configuration used during generation. The appendix proposes a response-based monitor that infers the applicable rule and selects a corresponding detector when the original prompt and generation metadata are unavailable. Because text-dependent detector selection requires procedure-level calibration and may not identify the rule actually used, the authors leave empirical evaluation of this setting unresolved.

References

Rule recovery, detector equivalence, and detection power are distinct outcomes, particularly after paraphrasing; we leave their evaluation to future work.

— Constitution-Guided Watermarking  (2610.09552 - Aremu et al., 7 Oct 2026) in Appendix, Section “Framework Extensions,” subsection “Prompt-Free Detector Selection”

We hypothesize that training them for their roles would improve performance: a verification agent supervised on watermark implementations, checked derivations, and empirical failure cases may propose stronger initial configurations and need fewer rounds, and a monitor trained on multi-rule labels may route more reliably, which Table~\ref{tab:router-comparison} shows directly affects detection, quality, and compliance.

— Constitution-Guided Watermarking  (2610.09552 - Aremu et al., 7 Oct 2026) in Appendix, Section “Extended Discussion and Limitations,” subsection “Hypotheses for Improvement”