Trailing-segment fragile-score detection for edits near the text end

Develop a trailing-segment version of the fragile-score detector to identify tampering when edits are confined to the final tokens of Cocktail-watermarked text.

Background

Cocktail combines a robust watermark signal for provenance with a fragile signal for tamper evidence. The fragile signal is seeded over a long normalized-text window, so edits generally corrupt downstream fragile seeds and reduce the fragile detection score. However, edits near the end of a document affect only the tail of the fragile seeds and may therefore be harder to detect when the detector aggregates the fragile score over the entire text.

The paper identifies reading the fragile score over trailing segments as a natural extension, but does not develop or evaluate such a detector. This unresolved direction is intended to improve sensitivity to tampering confined to the final tokens while retaining provenance attribution through the robust signal.

References

Edits confined to the final tokens corrupt only the tail of the fragile seeds, and reading $f$ over trailing segments is a natural extension we leave to future work.

— Tracing Provenance and Detecting Tampering with Complementary LLM Watermarks  (2608.12713 - Feng et al., 13 Aug 2026) in Discussion and Conclusion