Evidence integrity and review utility in crop-image decision support: an audit-to-review framework for precision agriculture

Abstract Background Agricultural artificial intelligence benchmarks can overstate decision reliability when multiple files represent the same biological evidence or when split provenance is unclear. We evaluated a classifier-agnostic audit-to-review framework that fixes evidence identity and role provenance before model comparison, declares aggregation estimands, and links predictive uncertainty to review workload. A public wheat-leaf corpus of 7,595 files was hashed to 5,118 unique contents; eligibility criteria defined a 915-content five-class task. Archived file-level benchmarks were separated from three frozen near-duplicate-grouped P2 partition configurations, and semantic, texture and fusion representations were evaluated under partition-weighted and one-content-one-weight estimands. Split conformal prediction was assessed by empirical coverage, review rate, error capture, review yield and dimensionless review utility. Results Macro-F1 was 0.962 under the legacy P0 benchmark, 0.805 ± 0.030 for semantic-only P2 and 0.819 ± 0.029 for direct fusion. The descriptive P0-P2 workflow-sensitivity gap was approximately 15.7% points, versus a 1.4-point partition-weighted fusion increment (95% bootstrap interval − 0.006 to 0.036). The leading fusion-related variants were numerically close, so no inferential ranking is claimed. At α = 0.10, global split conformal prediction achieved 0.896 empirical coverage while reviewing 20.8% of images and capturing 51.4% of top-1 errors; class-conditional Mondrian calibration achieved 0.923 coverage while reviewing 29.7% and capturing 64.8% of errors. Conclusions Defining the evidence unit is part of the statistical specification of a crop-image benchmark, not merely a data-cleaning step. In this corpus, benchmark interpretation was much more sensitive to evidence-workflow specification than to the evaluated representation refinement. The framework supports auditable internal evaluation and review triage but does not establish field, farm-level or intervention performance.