Untargeted plant LC–MS/MS routinely produces far more reproducible spectral features than can be assigned to authenticated molecular structures. The resulting gap between detection and identification is increasingly filled by spectral-library search, molecular networking, formula inference, class prediction, candidate ranking, predicted spectra, retention modeling, and de novo structure generation. These tools expand chemical coverage, but their outputs are often reported at a level of structural certainty that exceeds the evidence actually available. This article develops an original computational analytical framework for preserving uncertainty across heterogeneous annotation routes rather than forcing every feature toward a single named compound. The analysis separates direct spectral evidence, analogue and neighborhood evidence, formula-level inference, class-level inference, ranked structural candidates, generated structural hypotheses, and orthogonal analytical constraints. It further identifies recurrent points at which confidence is lost, including ion-level redundancy, erroneous fragment interpretation, incomplete reference and candidate spaces, model-domain mismatch, chemically biased training data, and correlated prediction errors. The proposed scheme treats confidence as a structured evidence state rather than a single universal score and allows unresolved, class-supported, family-supported, candidate-ranked, or exact-identity outcomes to remain distinct. Its intended value is interpretive discipline: downstream plant metabolomics conclusions should inherit the uncertainty of the annotations on which they depend. The framework is not prospectively validated, does not establish universal numerical thresholds, and requires benchmarking across instruments, laboratories, chemical domains, and candidate-space conditions before operational use.
