How a finding is decided

A coverage gap is decided by a rule, not by a model. Each rule is a fixed condition on fields extracted from the document — “the document says earthquake is not covered”, “no beneficiary is recorded”. The same rule on the same fields always gives the same result.

The AI does two things: it reads the document and fills in the fields, and it writes a finding's explanation in plain language. It does not decide whether a gap exists or how serious it is — the schema we ask it for has no such field.

A rule's “severity” exists in the database because the schema requires it. It renders nowhere and orders nothing: findings are ordered only by the provenance of the requirement and by the catalogue's declared order.

The same document, with the same rule catalogue, yields the same findings. Every finding records which rule decided it, with which engine version, on which values, and from which analysis it came — dated.

When the catalogue grows, an older analysis does not silently acquire checks it never ran: the composition is always computed against that analysis's own check plan, and one sentence says checks were added since. A new analysis is your action, never automatic.

The catalogue holds 50 rules across 14 lines (Group Health, Group Life, Health, Home, Income Protection, Life, Motor, Motorbike, Pension & Savings, Personal Accident, Pet, Renters, Roadside Assistance, Travel) out of the 32 lines the product accepts. The remaining 18 lines have no check yet: there the app says “we have not assessed this” and shows neither a composition nor a reassurance.

Every rule reads fields the extractor actually fills for its line, declares its inputs up front, and ships with two test documents: one that fires it and one where it correctly stays quiet. Without those, a rule does not enter the catalogue.

Checks that ask “is this recorded?” fire on the document's silence and always say “not recorded”, never “not covered”. Coverage checks fire only when the document explicitly says something is not covered. Whatever was not read stays “could not be checked”.

Every rule carries a classification of the requirement behind it: legislative (a law and article), contractual (a named class of contract), market practice (a named public source) or under review. Today 8 requirements are classified with a citation and 42 remain under review.

A classified requirement shows its citation beside the finding, everywhere. Requirements under review render in their own section, count in no summary, and reach no notification, email or report. The classification is the authors' proposal; the legal and underwriting reviews sign it before general availability.

We do not give a protection score, an index out of a hundred, or a severity colour. A number out of a total appears only with its denominator and what was counted.

The absence of a finding is not reassurance. Three states say explicitly what was not checked: an analysis that predates the check plan, a line with no checks, and a catalogue that changed after the analysis.

We do not compare products, suggest switching insurer or give advice. Findings are a prompt for a conversation with your adviser — who sees the same object you see.