Provider manuals remove judgment by moving the rule elsewhere
One chapter of a provider manual from an Illinois Medicaid managed care plan, encoded against 42 CFR Parts 447 and 438.
What this page establishes, and what it does not
A finding about published documents. Not a claim about any institution's conduct, and not a claim that any determination was wrong. Every figure is over one chapter of one manual, an archived revision, and the regulations as served on 20 August 2026. No manual's clause text is reproduced. Nothing generalises beyond the documents measured.
Elsewhere-heavy, not judgment-heavy
We predicted, before decomposing, that a manual written to remove judgment from the provider would be mostly mechanical. It was not. Of the 213 units in the chapter, 110 could be expressed as a testable clause and 103 could not, and the largest class among those that could not is incorporation: 22 units whose content is another document. Of the 110 expressible units, 14 are mechanical and 9 require judgment.
The manual removes judgment by moving rules out of itself, not by stating them. That is why conversion is a pipeline with fixed inputs and outputs rather than judgment work, and a more useful finding than any divergence figure.
The count depends on the split. One unit per heading instead of per sentence, 9 of 24 units read as unexpressible, against 103 of 213. The finer figure is reported.
No two governing documents agree on what a clean claim is
The regulation defines the term. 42 CFR 447.45(b), paraphrased: a clean claim can be processed without further information from the provider or a third party, excluding a claim from a provider under fraud or abuse investigation and a claim under medical necessity review. 447.46(b) gives managed care organisations that meaning, and 447.46(c) carries the timely-payment rules of 447.45(d) into their contracts.
Provider manuals from Illinois Medicaid managed care plans start their processing clock on that term, and no two governing documents agree on it. The term is defined with elements the regulation does not have and without its fraud exclusion, or used with no definition and no citation to 447.45 or 447.46. No reader of any single document can see the disagreement.
A quarter of the disagreement is a manual unable to decide
6 paired rules were evaluated against their regulatory counterparts over 4,000 generated fact sets each, 24,000 per side, by one interpreter. With the manual's term left undefined, the sides disagree on 15,536 of the 24,000 generated fact sets.
Of those 15,536 divergences, 3,904 (25.1%) are cases where the manual cannot determine at all, because its operative term is defined only in a regulation it does not cite. Supplying the regulation's definition removes that class: of the 3,904, 2,659 become operative differences a reader of both documents would catch, and 1,245 become agreements.
The rest of the 15,536: 6,040 (38.9%) differences a reader of both documents would catch; 4,000 (25.7%) the regulation defers a parameter the manual states; 1,592 (10.2%) the manual determines where the regulation leaves a judgment unmade. Shares are of sampled fact sets, not rates of real claims.
What was not done
Nobody answered for the institution. The 14 ambiguities and 9 undefined terms raised stand unresolved. The manual was an archived revision, chosen for tractability. Conversion cost is not measured. One programme is no population.
Check it
42 CFR Part 447 as measured: eCFR, 20 August 2026, sha256
01ac3dae19141563320321aac4e4fcfd4afa05760870dd154875aec3eb1546aa. The
interpreter, the register schema and the router are in
observer-protocol/op-policy-engine.
The manual's register is not published.