AI Med Policy Research

A Codex skill that turns AI-assisted medical & pharmaceutical policy research into a traceable, gated agent workflow — methodology isolation, RAW-data governance, PROSPEC packets, a hard state machine, and a validator.

AI Med Policy Research is a Codex skill for AI-assisted medical and pharmaceutical policy research. It turns policy-research work into a traceable agent workflow with prospective methodology design, RAW-data governance, structured-data protocols, PROSPEC execution packets, hard state-machine gates, and evidence-bound quality review.

It’s built for workflows where an agent may extract a methodology from literature, hand it to sub-agents, run analysis against raw data, and then compare outputs — without leaking known conclusions or paper identity into execution.

What it solves

AI agents produce useful policy drafts but fail in ways that are hard to audit: skipping data discovery, letting known conclusions shape the method or the review bar, leaking paper titles / DOIs / filenames to the execution agent, claiming a review passed without evidence, or patching a flawed database instead of rebuilding from governed sources. This skill makes those failure modes explicit and blocks completion until the required artifacts and validator checks support the claim.

Core capabilities

  • Research contract — pin down the policy object, question, boundary, expected output, data scope, contamination boundary, and minimum acceptable output.
  • Methodology & source-identity blinding — keep methods separate from known results, and keep paper title / DOI / URL / filename out of execution-facing artifacts.
  • RAW-data gate — require raw_inventory.json + a synchronized README_DATA.md before any analysis, cleaning, ingestion, or modeling.
  • PROSPEC execution — machine-readable packets carrying prospective expectations, procedure, evidence trace, and a completion gate for every subtask.
  • State machine — INIT → METHOD_LOCKED → TASKS_LOCKED → EVIDENCE_COMPLETE → REVIEWED → ACCEPTED, with ITERATE for failed gates.
  • Validator-gated completion — a deterministic Python validator rejects skipped transitions, missing artifacts, incomplete RAW coverage, public-identity leaks, and unsupported ACCEPTED claims.

It’s a workflow-control and research-governance skill — not a substitute for clinical, legal, or regulatory judgment.

Stack: Codex skill (SKILL.md + reference protocols + a Python validator) · bilingual English / 中文

Repository: wcboy/ai-med-policy-research