Design an identifiable ablation of the Critic and decomposition mode

Construct an experimental design in which disabling the Critic or changing the atomic-claim decomposition mode can alter the committed answer, thereby isolating the causal effects of those components on GANDR’s performance.

Background

The GANDR orchestrator commits a Round-1 draft whenever the structural protocol check passes, without consulting the Critic’s verdict. Consequently, disabling the Critic or changing the decomposition mode generally leaves the committed text unchanged, making their apparent accuracy effects unidentifiable apart from sampling variation. The authors report that a different design is needed to make these components causally testable.

References

Isolating the Critic and the decomposition mode would require a design in which they can change the committed text, which we leave to future work.

— GANDR: Claim Auditing for Verifiable Legal Answer Generation  (2609.10293 - Qian et al., 9 Sep 2026) in Appendix B, Section “Why only one ablation is reported”

This paper describes a system and does not evaluate it in a controlled way. We have not yet measured recall against human-curated reviews, the precision of extracted claims, or the contribution of each component, such as decomposition versus a single query or snowballing versus none.

— Optimist Deep Research: An Iterative, State-Centric Agent for Autonomous Literature Research  (2610.23299 - Panapitiya, 16 Oct 2026) in Section 11, Discussion and Limitations, paragraph “Evaluation”