Proxi GUIDE manual
Browse
On this page

Project internals

Agent review protocol

How bounded reviewers return findings without rewriting shared source material.

Project methodSource: 10_Agent_Review_Protocol.md

Status: Governing workflow, v0.1

Purpose#

Allow specialist agents to review individual SOPs without reloading the complete conversation history or silently changing settled architecture.

Minimum context for every review#

Provide the reviewing agent:

  1. README.md
  2. 01_Architecture_and_Control_Principles.md
  3. 02_Roles_and_Escalation.md
  4. The target SOP
  5. 06_Edge_Case_Register.md
  6. 08_Decision_Log.md
  7. Relevant sources from 09_Source_Register.md

Required review output#

Agents must return:

  • Findings ranked as critical, material, or minor.
  • Exact affected section/rule/edge-case IDs.
  • Missing edge cases.
  • Unsafe automation or unnecessary human work.
  • Contradictions with governing architecture or CMS source.
  • Proposed edits as discrete diffs.
  • New open decisions rather than assumed answers.
  • Test cases that would have caught each issue.

Prohibited review behavior#

  • Rewriting the complete SOP without explaining changes.
  • Changing a SETTLED decision without a change proposal.
  • Treating an LLM classification as authority.
  • Adding a human review merely because the agent is uncomfortable with automation; it must identify the unresolved fact or authority requirement.
  • Automating an action merely because it is technically possible; it must identify the deterministic rule and evidence.
  • Conflating public CMS guidance with the controlling executed Participation Agreement.

Integration discipline#

  1. Review findings are recorded before edits.
  2. Accepted changes update the target SOP and, where applicable, the edge-case register and rule catalog.
  3. Material design changes receive a decision-log entry.
  4. Rule IDs are stable; semantic changes increment rule versions.
  5. Every accepted edge-case change receives at least one test case.
  6. The project index/changelog is updated after integration.

Suggested agent assignments#

  • CMS/GUIDE compliance reviewer
  • Clinical dementia-care reviewer
  • Medication-safety/pharmacy reviewer
  • Care-navigation operations reviewer
  • Workflow/state-machine engineer
  • Security/privacy/interoperability reviewer
  • Billing/audit-evidence reviewer
  • Adversarial edge-case reviewer
  • Staffing/cost-model reviewer