Where an additive field correction stops

Trust requires matching evidence and action granularity to the structure of model failure. This page separates three questions: what information reaches the model, what residual structure remains, and what an additive field-constant correction can change.

Spatial reliability gap

additive field action versus sub-field residual structure

Moran's I spatial autocorrelation for labels, predictions, and residuals

Information and action granularity across the pipeline

context only, not a causal diagnosis

Real Sentinel-2, terrain, soil, and weather inputs leading to pixel prediction, field evidence aggregation, a scalar action, and spatial audit

Input supportSatellite, terrain, and soil are sampled at the query pixel; weather is one changing daily series shared within the field.

Evidence boundaryARC uses a nominal 349-feature agronomic representation once per field before selecting historical analogues.

Action boundaryThe evaluated action broadcasts one additive offset to every pixel, so it changes field mean bias but not centered residual geometry.

Interpretation boundary. Native support is shown to describe the data pipeline. It does not establish that modality resolution caused the residual clusters or that finer inputs would necessarily remove them.

An additive field-constant action cannot move local clusters

current F1.4 showcase · wheat LORO · SR 0.2

Observed yield, frozen residual clusters, and ARC residuals for the current wheat LORO showcase field — **Negative-result boundary.** The centered frozen-model error mask is reused on the ARC error panel; LISA is not recomputed after correction. ARC changes the field-level error magnitude but does not claim to remove the within-field pattern. Open PDF.

ARC changes field-level error magnitude while the centered within-field pattern is held fixed. The displayed mask is defined once from frozen prediction error and reused for comparison. underpredictionoverprediction

Method details

Error follows the field explorer convention, prediction − GT. The frozen-derived mask uses 8-neighbour queen weights, 999 permutations, Benjamini–Hochberg FDR q≤0.05, and a retained-pixel share of about 4.5%. The same mask is applied to the ARC error panel; LISA is deliberately not recomputed after correction. This is an explanation-alignment boundary and a negative result, not an agronomic or measurement diagnosis.

The continuous additive oracle quantifies the action-class floor

six crop × protocol settings

Field-balanced decomposition of residual sum of squares into removable field-mean bias and the within-field additive-scalar floor

Within-field floor28.1–63.1%

Share of field-balanced sampled-pixel residual SSE left after the optimal continuous field offset.

Coverageppf256

Exact for the available sampled OOF support, not an exact full-pixel population floor.

Claim scopeadditive only

The result does not cover spatially varying, multiplicative, affine, or management-zone actions.

Two oracles answer different questions. This continuous oracle isolates the within-field floor for any real-valued additive offset. The separate five-action oracle measures perfect routing headroom within the implemented policy menu.

Counter-intuitive evidence signals

helpful versus harmful evidence-card cases

Helpful and harmful evidence-card cases compared by deploy-visible signal strength

The harmful case looks more convincing on the obvious signals. Neighbour agreement and view consistency are both perfect in both cases, while the harmful case has a stronger prediction z-score, stronger anomaly score, closer nearest neighbour, and larger bank. Evidence disclosure is therefore valuable because it shows why forced action selection is unreliable.

Evidence Card v2 paired cases

same evidence machinery, opposite deployment outcomes

Fold landscape and target support versus fold gap