BlogBenchmarks6 min read
Reliability

Evidence before confidence

A confidence score without evidence is not enough. PolDex treats cited source text, page context, and truth state as part of the output contract.

Confidence is not proof

A model can be confident and still be wrong. In insurance extraction, the customer does not only need a number. They need the reason that number exists. That reason should be visible as source text, page context, form role, evidence quality, and the rule outcome that allowed the value to pass.

PolDex starts with evidence because downstream systems need auditability. A broker, MGA, carrier, compliance product, or agent workflow should be able to inspect the source that supported a field instead of trusting an opaque completion.

Evidence changes product behavior

When evidence is required, the extraction engine becomes more careful. It cannot simply fill every field with a plausible answer. It has to attach the value to a source span, normalize it into the schema contract, and preserve uncertainty when the support is not strong enough.

This is why FastScript returns truth states such as present, absent, unresolved, conflicting, low confidence, unsupported, or review required. The state is not a decoration. It tells a downstream system how to handle the field.

What developers get

The output is structured JSON, but the evidence travels with it. That lets developers build review screens, compliance checks, agent workflows, and exports without reverse-engineering the document again.

The long-term goal is simple: every value PolDex returns should have a reason to exist, and every public accuracy claim should point back to a named benchmark corpus rather than marketing language.