Skip to content
Applied AI Systems/Case studyIn Development

Healthcare policy intelligence: which revision applied, and when

An architecture for policy decisions where an apparently correct answer can still be wrong because it rests on the wrong revision, date or authority: a 2025 prototype that resolves the revision in force on the date, and a proposed modernization that reconstructs the source first and reasons over procedure behind deterministic gates.

role: Senior AI Architect (independent project)organization: DeNovo Innovations Groupperiod: 2025 – presentdomain: Federal health-benefit policy manuals with long revision histories

Labels on this page: VERIFIED — PUBLIC · VERIFIED — PRIVATE · SECONDARY · WITHHELD.

Ecosystem · Federal rules, national coverage policy, regional contractor adoption, state variation; each with its own revisions and effective dates

Problem

A question about coverage can be answered identically across many revisions of a manual and differently in the one that applies. Similarity search ranks the many and buries the one. Applicability also depends on effective date, authority relationships and the revision history of the source.

Covered? Authorized? Payable? Each answer is for this patient, this provider, this place and this date of service. The work is reconstructing the applicable rule and its context, not finding a passage that reads well.

Figure 1 · The information problem

SECONDARY
Ecosystem map: a federal health agency governs four manuals, Policy, Operations, Reimbursement and Systems, each with its own revision history; national payer rules, state and locality variation, regional contractors and coverage exclusions all feed one decision: covered, authorized, payable, for this patient, provider, place and date of service.
One question, an entire ecosystem. Four manuals with four revision histories, governed centrally, implemented by regional contractors, cross-referenced with national payer rules and state variation. An overview of the problem space; revision tracks are illustrative.sha256 7e93fa91a5698605… · Contractor arrangements and manual editions as of Oct 2026.

Architecture

The paper sets out the full design and its commercial vision: where policy intelligence turns into operational value, and how the system stays current as policies change.

Policy intelligence is two questions that cannot be separated: which rule applies, and what that rule requires (the actions, conditions, exceptions and timing it imposes). My architecture answers them in a fixed order: reconstruct the source first, then reason over it, then let only verified output through.

LayerStageStatus
Source truthExtraction of sections, tables, images and cross-references; before-and-after reconstruction from change markupBuilt (2025 prototype)
Source truthRevision graph over a manual's full release history, with content inheritanceBuilt (2025 prototype)
Source truthPinned releases, an element ledger and a checked canonical reconstructionProposed (Gate A)
ContextResolve the revision in force on the date, then run the vector search inside itBuilt; exercised on one section
ContextEffective, publication and observation dates; supersessionProposed (Gate B)
ContextGraph and vector search in one store, with the date filter and the similarity search in one queryProposed
ContextSelf, parent, ancestor and subtree views of each sectionProposed
ProcedureTyped procedure and step objects: actors, triggers, deadlines, forms, exceptions, appealsProposed (Gate C)
ProcedureOntology seeded from the manuals' own definitions and acronyms, with is-a and part-of relationsProposed (Gate B)
Verified outputA directed workflow of specialist roles and small models behind deterministic gatesProposed (Gate D)
Verified outputExplicit UNKNOWN or PARTIAL with human review; reproducible as-of answersProposed (Gate D)
ScopeNational coverage policy and regional contractorsAfter the gates pass

Decisions

  • Canonical first. The reconstructed source and its procedure objects are the record of authority; embeddings and model output only interpret it.
  • Resolve the date, then search. The revision in force on the date is resolved before any similarity search, so that a passage from the wrong revision cannot reach the answer. In the 2025 prototype the vector search followed this rule; context expansion and undated questions did not yet, and the modernization closes both, running the date filter and the search in a single graph query.
  • Correct means in force on the date. An answer counts as correct only if it cites the revision that governed on the date asked.
  • Procedures as structure. Actors, steps, deadlines, forms and exceptions become typed, cited objects, because instructions read as flat text lose their order and conditions.
  • Models propose, gates decide. Small models propose extractions and mappings; deterministic checks confirm source, scope and date before anything is used.
  • Abstain explicitly. When sources disagree or steps are missing, the system returns UNKNOWN or PARTIAL and routes the question to a person.

What was built

  • A working prototype on one manual section across its revisions: extraction of changes, before-and-after reconstruction, tables and links
  • A revision index over a manual's full release history, with each section linked to the release that last changed it
  • Date-based revision resolution: the prototype selects the latest revision released on or before the date asked and restricts retrieval to it (by release date; effective dates and supersession were designed but not populated)
  • Graph context (Neo4j) and vector retrieval (Weaviate) coupled to source citations, answered by a small local model
  • Revision identifiers and release dates carried through ingestion, the graph, the vector index and the citations

Proposed modernization

  • Canonical reconstruction, effective and observation dates, supersession and an ontology, in sequenced gates with acceptance evidence for each
  • Procedure extraction and a controlled workflow of specialist roles and small models, behind deterministic checks
  • Graph and vector search consolidated in Neo4j's native vector index, so the date filter and the similarity search run in one query

Verification

The 2025 prototype ran its date-resolution path end to end on real queries; it was not formally evaluated. Planned: evaluation sets of dated questions, with canaries that must fail; an answer counts as correct only if it cites the revision that governed on the date asked. The question-and-answer sets are being built during extraction, with a human in the loop checking accuracy.

Evidence

  • VERIFIED — PUBLIC
    Public account of the architecture (Global Trial Accelerators blog, Oct 2026)
    source · checked 2026-10-27
  • VERIFIED — PRIVATE
    Repository and technical whitepaper (2025)private
    held
  • VERIFIED — PRIVATE
    Evidence audit of the 2025 prototype (Oct 2026)held
    held
  • VERIFIED — PUBLIC
    Public architecture white paper: Health Policy Insight Engine (Nov 2026, PDF)sha256 323265c1439721cd1ee41c365661ad2b05993f78dfe027520041e04ec84e23ac
    PDF · checked 2026-11-07

Status

In development: a 2025 prototype, and a modernization that is proposed, not deployed.

Independent project: a 2025 prototype on one manual section; the modernization is proposed, not deployed. · updated 2026-11-07