BioTraceGuard A Provenance Gated Verification 2 Layer for Agentic Biological Data Analysis

Tool-using agents can assemble increasingly complex biological analyses, but a plausible plan and a successfully executed pipeline do not guarantee that the resulting scientific claim is statistically or biologically admissible. We introduce BioTraceGuard, a model-agnostic execution layer that separates agent planning from claim authorization. The planner may choose tools freely, while a typed provenance ledger records data identities, split membership, transformation scope, target access, statistical family, domain structure, uncertainty, and counterfactual checks. Claims are released only when the trace satisfies seven gates covering provenance completeness, partition isolation, target leakage, task-family constraints, null-signal testing, uncertainty, and claim stability. We validate the design on five public biomedical datasets spanning classification, regression, count data with exposure, stratified case-control tables, and right-censored survival analysis. On Wisconsin Diagnostic Breast Cancer, a valid pipeline achieves mean AUROC 0.9948 ± 0.0049, while an intentionally target-leaky pipeline reaches 1.000; the guard rejects the latter. On the diabetes progression data, valid repeated-CV R 2 is 0.4853 ± 0.0699 versus 0.9984 under target leakage, again rejected. A count-analysis gate detects Poisson overdispersion of 2.57 and escalates to a negative-binomial model; censoring and stratification gates route transplant survival and eight-city smoking data to Cox and Mantel-Haenszel analyses. In 1,500 executable conformance tests, all 1,000 injected invalid traces are rejected and all 500 valid traces are accepted; 500 benign schema perturbations leave routing unchanged. These tests are intentionally white-box and do not substitute for naturalistic LLM-agent evaluation. They instead establish a reproducible safety primitive for scientific agents: high apparent performance cannot authorize a biological claim without a valid computational lineage