Extraction Hallucination

Explanation

Extraction Hallucination checks that a document-extraction system reports only values that genuinely appear in the source document and never fabricates field values — critical in BFSI where automated extraction pulls figures such as account numbers, income, or balances from statements and KYC documents, and an invented value can flow straight into a credit or onboarding decision. It is measured by the Extraction Hallucination Rate (metric and calcMethod are identically named): the proportion of extracted fields that are fabricated, i.e. not supported by the source text. Implement it by grounding every extracted field to a verifiable span or bounding box in the source, then validating extractions against that evidence — through automated source-span matching plus human review on a sampled or flagged subset — and log each field, its source location, confidence, and the fabricated/verified verdict so the evidence trail is complete. The threshold is < 0.1% fabricated fields, and this Immediate-priority control mandates human routing on breach: whenever the fabricated-field rate crosses 0.1%, affected extractions must be diverted to mandatory human review rather than auto-processed, halting straight-through automation until the cause is found and the extraction pipeline corrected.