Agronomic decision support never has a single source of truth. The crop model says one thing, the satellite another, the soil-moisture probe a third, and the operations log records an irrigation that the data “doesn’t see”. Most systems either look at each source in isolation or return a final recommendation without saying what conflicted.
In our paper, published in the SETN 2026 proceedings (ACM), we organise all sources into a shared contextual state of the field. Over that state the system:
- detects four contradiction types: cross-source value, action–observation, temporal drift and model–reality;
- weights the evidence deterministically to compute a consensus value;
- explains the outcome in text and, where it helps, tables and plots, narrated by a large language model.
Across five smart-agriculture scenarios (cotton and wheat, Orchomenos, Greece), state-aware explanations beat text-only baselines on contradiction detection (accuracy 0.92 vs 0.66 and 0.86), coverage, outlier-source attribution (0.96 vs 0.27 and 0.86) and actionability. Three agronomists rated the explanations 4.3–4.6 out of 5.
This research shapes how our tools evolve, so that AI tells you not only what to do but why, even when the data disagree.