sanity controls and selector validation

Controls for correction governance

These diagnostics test whether q_agro evidence supports deployment decisions rather than only post-hoc narratives. The page is generated from the same artifact summary used by the manuscript spec.

Control summary

S1-S5 plus P1-P5 selector validation

S1 Risk-coverage abstention

rescued settings against random-ranking nulls

Risk coverage curves for rescued q_agro correction settings — **Selective abstention.** Curves show settings where all-qagro is harmful but correcting only the most reliable coverage becomes beneficial. Null controls use random rankings.

Crop	Protocol	Stage	Score	Coverage	All q_agro gain	Selective gain	Random p95	p-value

P5 Evidence scores

helpful-correction AUC

Score	AUC	Settings AUC > 0.5	Visual

P2 Universal policies

fixed policy generalization

Policy	Gain vs base	Gain vs all q_agro	Positive settings	Beats all q_agro

S3 Rolling-history nulls

real signal versus randomized controls

Crop	Protocol	Real beats nulls	Mean real gain	Real minus best null

S4 Spatial gap

Moran's I and residual structure

Crop	Protocol	Label I	Prediction I	Residual I	Residual significant fields