Method
1. Ingest. The whole policy manual or handbook, split into sections with code, title and revision date.
2. Evidence bank. 57 policy domains. Each has the research finding, its effect size, an evidence tier (A: meta-analysis or several causal studies; B: one strong causal study; C: correlational; D: professional standard), the counter-evidence, the testable standard, a model clause, and red-flag language. Every citation is resolved against Crossref or DataCite and its abstract pulled from OpenAlex before it is shown.
3. Reading. A calibrated decision model (Jev) reads every section against every domain and every stated goal: does it govern the practice, does it meet the standard, does it conflict, is the commitment measurable, is it stale. A conflict has to be visible in the section's own words; a missing piece is scored partial, never conflict. Long sections are read at the opening and around every passage that touches the domain, so a prohibition on page three is not cut off. A human reviewer can override any verdict, and the override is printed on the report. About fourteen thousand judgments per manual, for well under a dollar.
4. Score. Evidence alignment 40%, goal alignment 20%, specificity 15%, currency 10%, coverage 15%. The formula is printed on every report.
5. Recommendations. For every gap, the district's own sentence, the research, and the model clause rewritten in the district's numbering. Every number in a rewritten clause is checked against the inputs.
6. Implementation and success. Every domain carries a plan (first 90 days, year one, ongoing, with an owner) and a success test (the measure, how the baseline is set, the target and its timeframe, the research benchmark, and when it is reported). A district that adopts the manual gets a one-page board dashboard of those measures. Success is not "the policy passed"; it is the measure moving by the date the policy names.
7. Corrections. Anyone can dispute a verdict or submit evidence from any domain on any report. A person reviews each one; accepted corrections change the output and are listed with what changed on the corrections page.
Honesty rules. Nothing is shown with an unverified citation. Tiers and counter-evidence are printed. Where state law settles a domain, silence is not penalized. Every verdict shows its receipt so a reader can disagree with it.