Renderer, not source of truth
A model can draft a clear explanation, but the finding must be grounded in structured evidence, source URLs, scan observations, confidence, and validation requirements.
Required inputs
Evidence payload
The source page, structured data field, policy excerpt, channel response, AI observation, or connected merchant source used for the finding.
Finding type
The issue family and merchant-facing consequence that determine the shape of the explanation.
Confidence and validation
The confidence label and the retest path that bound how direct the recommendation can be.
Repair and fallback
Generated copy that drops evidence, overstates certainty, invents outcomes, or loses the validation path should be repaired or replaced by deterministic copy.
Private boundary
StoreSteady publishes the governance principles and public policy shape. It does not publish internal prompts, full evals, exact thresholds, scoring weights, repair logic, adversarial tests, or vendor-specific prompt details.