RISED: A Pre-Deployment Evaluation Framework for High-Stakes AI Decision-Support Systems, with Application to Healthcare

\textbf{Background:} Regulatory frameworks for AI in healthcare, including the EU AI Act and FDA guidance on AI/ML-based medical devices, require clinical decision support to demonstrate not only accuracy but auditability. Existing formal languages for clinical logic validate syntactic and structural correctness but not whether decision rules use epistemologically appropriate evidence. \textbf{Methods:} Drawing on design-by-contract principles, we introduce meta-predicates -- predicates about predicates -- for asserting epistemological constraints on clinical decision rules expressed in a DSL. An epistemological type system classifies annotations along four dimensions: purpose, knowledge domain, scale, and method of acquisition. Meta-predicates assert which evidence types are permissible in any given rule. The framework is instantiated in AnFiSA, an open-source platform for genetic variant curation, and demonstrated using the Brigham Genomics Medicine protocol on 5.6 million variants from the Genome in a Bottle benchmark. \textbf{Results:} Decision trees used in variant interpretation can be reformulated as unate cascades, enabling per-variant audit trails that identify which rule classified each variant and why. Meta-predicate validation catches epistemological errors before deployment, whether rules are human-written or AI-generated. The approach complements post-hoc methods such as LIME and SHAP: where explanation reveals what evidence was used after the fact, meta-predicates constrain what evidence may be used before deployment, while preserving human readability. \textbf{Conclusions:} Meta-predicate validation is a step toward demonstrating not only that decisions are accurate but that they rest on appropriate evidence in ways that can be independently audited. While demonstrated in genomics, the approach generalises to any domain requiring auditable decision logic.

Benchmarks are necessary for healthcare evaluation, but are not sufficient for predicting deployment performance. Our position is that the evaluation--deployment gap arises not because of poorly designed benchmarks, but from implicit assumptions about how users interact with models that cannot be surfaced from benchmarks alone. To make this precise, we propose a classification of assumptions into two categories: task, which can be tested from conversation data alone, and outcome, which requires outcome data and behavioral studies for testing. Critically, outcome assumptions depend on human behavior, something that even well-designed benchmarks cannot directly observe. To demonstrate the operationality of this framework, we retrospectively analyze a healthcare RCT as a case study and find that the gap naturally separates into task and outcome gaps of roughly equal size. To address this, we make two contributions: first, we propose BenchmarkCards, an artifact that documents assumptions, and second, we propose staged evaluation, a procedure that systematically tests assumptions and evaluates performance.

Two clinical AI systems can score nearly identically on coverage-based rubrics yet behave radically differently when their patient inputs change: one updates its recommendations to match the new clinical signal, while the other produces the same output regardless. We introduce the Causal Sensitivity Score (CSS), a pre-registered interventional metric that mutates oncology tumor-board cases along five clinically meaningful dimensions - biomarker flips, prior-treatment failures, biomarker removals, surgery-status changes, and stage perturbations - and scores whether each model updates its recommendations in the pre-registered correct direction using a {0, 0.5, 1.0} scale. Benchmarked against the Consensus Match Score (CMS), a coverage-based weighted recall metric, six frontier models from three labs evaluated in single-shot inference across 224 cases rank in nearly opposite orders: all six models change rank, the CMS-worst model becomes CSS-best, and one upper-mid CMS model ranks last on CSS. We further surface a universal safety blind spot: every frontier model fails on surgery-status interventions (at most 17.2% CSS on Family D), a finding CMS does not expose. The metric also transfers to tool-using agents: in a ReAct-style experiment, tool use improves CSS for five of six models (+2.5 to +20.3 percentage points), yet the lowest-CSS model retrieves the same chart sections and still fails to update its recommendations - revealing a structural responsiveness deficit visible only under counterfactual evaluation. Cross-judge replication and three-rater medical-professional validation confirm the aggregate findings. Interventional pre-registered metrics like CSS complement coverage-based evaluation for clinical AI agents: they capture responsiveness that coverage metrics miss and offer a candidate dense reward signal for future agentic RL systems.