There is no standard for verifying AI citations. Current approaches use another AI model to check output, trading one source of error for another. There is no method to prove that a cited value was correctly extracted from a specific location in a specific document, rather than inferred, approximated, or fabricated.
Tasset's H Score changes that. Every line and cell in a Tasset document has an ADAP coordinate. When an AI cites a source from an ADAP formatted file, the H Score runs two deterministic checks.
F1: Does the anchor actually exist? If not, the citation is fabricated, caught instantly by a manifest lookup, no AI required.
F2: Does the quoted text match what is stored at that coordinate, character for character? If an AI drops words from a quote it may not be caught by a human reviewer. The H Score catches it because the string comparison has the stored text to check against.
Every citation gets classified and weighted. For fabricated anchors, F1= 1.0. For verbatim mismatches: F2=0.9. The weighted score across all citations produces a single number.
When every citation is verified accurate: H = 0.0. Every citation must be verified. No exceptions. Any score above zero flags which anchor failed, what type of failure, and what the stored text says versus what the AI quoted. H = 0.0 is the only clean result. Every fabrication leaves a structural trace. Every misquote leaves a structural trace. The score is math, not confidence.