ResistIA: Reasoning-Guided Agentic Evaluation of Synthetic Metal-Resistance Genes from Conditional Genomic Foundation Models

Genomic foundation models can generate plausible coding DNA, but turning raw generations into scientifically useful candidates requires reliable post hoc evaluation and robust model-selection criteria. We present ResistIA, a reasoning-guided agentic evaluation and refinement framework for synthetic metal-resistance genes generated from a LoRA-adapted GenomeOcean-500M model. ResistIA combines three complementary evaluators---(i) DIAMOND blastx protein homology, (ii) ORF/translation/physicochemical checks, and (iii) ChatNT-based semantic/regulatory Q\&A with embedding similarity---under an auditable decision policy that selects tools based on runtime availability and evidence needs. We study three progressively stronger versions of the framework: V1 (baseline multi-tool evaluation), V2 (closed-loop reweighting over generated batches), and V3 (validation-aware closed-loop optimization with score smoothing and early stopping). Across representative and multi-seed benchmarks, V3 improves reliability of iteration selection and shows stable validation behavior (best validation composite score $0.2505\pm0.0054$, 95\% CI: $[0.2458,0.2553]$) with high ORF success ($0.9662\pm0.0058$, 95\% CI: $[0.9611,0.9713]$). Validation DIAMOND pass rate remains lower and more variable ($0.2217\pm0.0459$, 95\% CI: $[0.1815,0.2619]$), highlighting homology robustness as the main source of run-to-run variation. In paired 10-seed fixed-vs-adaptive comparisons, we do not observe robust adaptive superiority, motivating a conservative reasoning-agent framing. These results position ResistIA as a compact, reproducible testbed for reasoning over heterogeneous biological evidence in agentic scientific workflows, and a practical framework for evaluator-guided refinement of genomic sequence generators.