What Do We Care About in Bandits with Noncompliance? BRACE: Bandits with Recommendations, Abstention, and Certified Effects

Researcher Nicolás Della Penna has introduced BRACE (Bandits with Recommendations, Abstention, and Certified Effects), a novel algorithm addressing a fundamental problem in AI decision-making: what happens when there's a gap between what an AI system recommends and what treatment actually gets delivered to users? This "noncompliance" scenario is common in real-world applications like healthcare recommendations, financial advice platforms, and content moderation systems where downstream actors (doctors, advisors, moderators) may override or modify AI suggestions based on private information.

BRACE is a parameter-free phase-doubling algorithm designed specifically for finite-context square-IV problems. The algorithm's key innovation is that it performs instrumental variable (IV) inversion only after matrix certification, otherwise returning full-range but honest structural intervals. This approach delivers three crucial guarantees: simultaneous policy-value validity, fixed-gap identification of the operationally optimal recommendation policy, and fixed-gap identification of the structurally optimal treatment policy under contextual homogeneity and invertibility conditions.

The research formalizes the objective-choice problem that platforms face—whether to optimize for recommendation welfare in current mediated workflows, treatment learning for future direct-control regimes, or anytime-valid uncertainty quantification. The paper demonstrates through examples that recommendation welfare can actually exceed every learner-measurable treatment policy when downstream actors leverage private information, challenging conventional wisdom about always optimizing for direct treatment effects.

Experimental benchmarks across five challenging scenarios show BRACE's practical value: safety manifests as regret on easy problems, as abstention and wide valid intervals under weak identification, as a reason to prefer recommendation welfare under homogeneity failure, and as tighter structural uncertainty when extra instruments are available. For rich-context problems, the research also derives an orthogonal score whose conditional bias factorizes into compliance-model and outcome-model errors, clarifying what must be stabilized for anytime-valid semiparametric IV inference.