What the documents actually say

Each indexed policy document is read back out of the index, split into criteria, and each criterion classified into the layer of a finding it governs — the coding rule, the clinical evidence, or the payment consequence. The result is then reconciled against the criteria the library already asserts for that version.

All policies Upload a policy

Concept, shading into pre-computed. On the platform this is AI Studio's decision_tree_extractor, an LLM component that posts its result to the Knowledge Center for UMLS enrichment and versioned storage. There is no model here — this is a deterministic structural parse: headings, sentence boundaries and a keyword-scored classifier. The shape of the output and the review loop around it are real. The extraction is not a model, and the screens say so wherever a number comes from it.

Criteria extracted

114

from 7 documents

Library criteria verified

supported by their own source document

Unverified

the extractor found no support — a reviewer decides

Needs a person

13 unclassified · 28 low confidence

Library criteria the extractor could not trace

1 of 21

A criterion an auditor applies that the extractor could not find in its own source document. That is a question to answer, not a verdict — the parser is not a model, and a near miss reads the same as a real gap until a person looks.

CPP-014 v4.2 CPP-014_em-level-selection_v4.2.md

A routine examination diagnosis without a documented problem addressed does not by itself support a level 4 or 5.

Criteria by layer

Across every indexed document

A determination that collapses these into one generic “unsupported” is what a provider disputes and what an appeal cannot answer, because it cannot tell which question was decided.

Coding rule 29

What the coding rule requires — level, sequencing, units, modifiers.

Clinical evidence 29

What the record has to show before the code stands.

Payment consequence 24

What it costs, and how the recoverable amount is computed.

Documentation required 9

What has to be on file, and how it is requested.

Provider dispute 10

How a provider challenges the finding.

Unclassified 13

No word in the classifier's vocabulary appeared. A person has to place these.

Every document

7 extracted

Document	Policy	Criteria	Coding	Clinical	Payment	Library traced	Review
CPP-014_em-level-selection_v4.2.md extracted 18 Sep 2026 18:47 UTC	CPP-014 v4.2	15	4	4	4	2 of 3 1 unverified	Validated
CPP-021_duplicate-and-repeat-services_v2.3.md extracted 18 Sep 2026 18:47 UTC	CPP-021 v2.3	13	5	2	2	3 of 3	Validated
CPP-047_inpatient-drg-validation_v3.1.md extracted 18 Sep 2026 18:47 UTC	CPP-047 v3.1	18	1	11	3	3 of 3	Validated
CPP-061_physician-administered-drugs_v1.4.md extracted 18 Sep 2026 18:47 UTC	CPP-061 v1.4	13	8	2	1	2 of 2	Validated
CPP-061_physician-administered-drugs_v2.0.md extracted 18 Sep 2026 18:47 UTC	CPP-061 v2.0	24	8	4	6	4 of 4	Validated
CPP-080_eligibility-enrolment-retroactive-termination_v6.1.md extracted 18 Sep 2026 18:47 UTC	CPP-080 v6.1	17	1	5	7	3 of 3	Validated
CPP-085_prior-authorization-post-service-review_v3.4.md extracted 18 Sep 2026 18:47 UTC	CPP-085 v3.4	14	2	1	1	3 of 3	Validated

CPP-021 is the control. It is the one seeded document deliberately written in a different shape — no numbered sections, no bulleted criteria, prose throughout — because a parser tested only against documents written by the same hand is testing its own fixture. Its numbers are in this table beside the rest, whatever they say.