Human-Governed Multi-Agent AI for Biological Hypothesis Generation: An Evidence-Grounded Framework Using JUMP Cell Painting Profiles

High-dimensional cellular phenotypes can motivate biological hypotheses, but similarity alone cannot establish mechanism or causal validity. DISCOVERY-AGENT links exploratory JUMP Cell Painting evidence with evidence-grounded hypothesis generation, independent critique and bounded human-review routing. An initial 65-record CRISPR/ORF analysis across nine gene–modality tests found four FDR-significant CRISPR plate-level associations (AKT1, BRCA1, MTOR and TP53), none among evaluated ORF genes. A distinct later cross-source compound-identity benchmark used 614 morphological features and achieved AUROC 0.6142 on 1,250 held-out compounds with training-only preprocessing; this is identity discrimination, not mechanism-of-action classification. Three structured hypotheses underwent LLM critique and revision, while eleven gene-pair evidence assessments yielded no FDR-significant mechanistic finding (minimum q = 0.2037). Archived governance checks passed 1,021/1,021 predefined tests; nevertheless, the initial live-agent run contained two execution errors among five cases, with both cases responding on fresh retest. These findings demonstrate feasibility of auditable evidence processing and modest cross-source morphology signal, not validated autonomous discovery, prospective safety, or completed expert review. The work delineates practical requirements for independent biological validation and reproducible scientific agents.