Hi BioStars community, A bioinformatics workflow can execute successfully and still be scientifically invalid. For example, an AI coding agent can perform cell-level differential expression across experimental conditions and return hundreds of significant genes even when the design suffers from pseudoreplication. I am developing BioNexus, an open-source scientific reliability layer for bioinformatics workflows, to address this class of failure. GitHub: https://github.com/HERRY423/BioNexus The motivation is simple. Coding agents such as Codex, Claude Code, Cursor, and similar tools are becoming increasingly capable of writing and executing bioinformatics workflows. However, being able to execute an analysis is different from knowing whether the analysis is scientifically valid. For example, consider a request such as: "Run differential expression between these two single-cell clusters across conditions." An agent may generate a workflow that treats thousands of cells as independent biological replicates. The code may run successfully and the resulting p-values may look convincing, but the inference itself may be invalid because the statistical unit is wrong. BioNexus is intended to operate at this boundary. It is not a standalone AI scientist, autonomous agent, notebook environment, or workflow platform. Instead, it provides a scientific reliability layer around existing coding agents and bioinformatics tools. The basic workflow is: analysis request - preflight -existing analysis tools - audit - evidence assessment - claim verification Current components include: machine-readable biological capability contracts deterministic scientific preconditions and refusal rules fail-closed behavior when required evidence or canonical backends are unavailable explicit separation between canonical execution and heuristic or experimental execution scientific failure taxonomy for common bioinformatics errors Claim-Evidence Ledger for linking conclusions to supporting evidence BioFailureBench, a collection of deliberately flawed biological analysis scenarios The current scientific failure taxonomy covers problems such as: pseudoreplication unsupported cell-type annotation identifier namespace mismatch experimental confounding missing multiple-testing correction parameter instability spatial coordinate substitution silent backend substitution biological claim inflation The current flagship validation track focuses on three areas. Single-cell differential expression validity The aim is to detect inappropriate cell-level inference across biological conditions and require valid replicate structure before population-level conclusions are made. Cell-type annotation evidence The goal is not to build another annotation algorithm. Instead, BioNexus evaluates how strongly a proposed annotation is supported by evidence such as reference mapping, positive markers, negative markers, doublet risk, open-set behavior, ontology compatibility, and cross-method agreement. Spatial inference validity The goal is to test whether a spatial biological conclusion survives plausible alternative explanations, including segmentation leakage, transcript density, cell size, local density, batch or FOV effects, neighborhood radius choices, and permutation controls. One design principle is particularly important: A heuristic method should never silently masquerade as a canonical biological method. Experimental and frontier capabilities are therefore separated from stable capabilities, and heuristic fallbacks must be explicitly disclosed. BioNexus is currently in the v0.10 stabilization stage. I am therefore especially interested in critical feedback rather than requests for additional features. The most useful feedback would include: real examples of bioinformatics analyses that run successfully but are scientifically invalid missing failure modes in single-cell or spatial transcriptomics situations where BioNexus refusal rules may be too conservative public datasets suitable for evaluating scientific validity checks approaches for measuring false-positive refusals researchers interested in testing BioNexus retrospectively on existing notebooks or pipelines I would also be interested in discussion around a broader question: As coding agents become capable of executing increasingly complex bioinformatics workflows, should "knowing when not to compute" become a first-class scientific capability? Critical feedback, biological failure cases, and contributions are very welcome.