Why a property score you cannot explain is a liability

The short answer

A model output that cannot be traced to its inputs is fine for triage and dangerous as a decision record. When a decision is questioned, the question is never what the score was. It is what was known, when it was known, and where it came from.

The four moments a file gets read by someone else

Underwriting files are written once and read under pressure. Four readers matter.

• The rate filing. A regulator reviewing a filing wants the basis for a risk classification, and increasingly wants to know whether a model produced it.
• The reinsurance treaty review. A reinsurer pricing your book wants to see the selection discipline behind it, not an assertion that it exists.
• The market conduct exam. An examiner sampling declines and surcharges wants the reason each one happened, reconstructed from the file.
• The disputed claim. Counsel wants to know what the carrier knew at bind, and a file that cannot answer that answers it badly.

What evidence-first actually requires

Three properties, none of which are exotic. Each fact records its provider. Each fact records when it was retrieved, because a hazard layer that was accurate at bind may not be accurate at renewal. And the raw provider response is retained, so a disagreement about what a source said is settled by reading the source rather than by argument.

This is not an argument against models

Imagery models that grade roof condition solve a real problem that no public dataset solves. The argument is about ordering. Assemble and cite the record first, then layer inference on top of it, so that when the inference is challenged there is something underneath it that is not also an inference.