Data-to-PolicyData-to-Policy

Policy teams receive reports far longer than anyone has time to read.Policy teams receive reports far longer than anyone has time to read.

Data-to-Policy turns a long socio-economic report into a two-page brief a decision-maker can act on. The hard part is trust: single-shot prompts drop statistics or invent them. So the work is split across agents with narrow jobs, the facts are schema-checked, and a reviewer can send drafts back until they hold up.Data-to-Policy turns a long socio-economic report into a two-page brief a decision-maker can act on. The hard part is trust: single-shot prompts drop statistics or invent them. So the work is split across agents with narrow jobs, the facts are schema-checked, and a reviewer can send drafts back until they hold up.

1. A graph, not a prompt. LangGraph models the work as a state machine, so every brief passes the same steps in the same order.

   A graph, not a prompt. LangGraph models the work as a state machine, so every brief passes the same steps in the same order.
2. Facts are typed. The extractor must return exactly five statistics that validate against a Pydantic schema, or the run fails loudly.

   Facts are typed. The extractor must return exactly five statistics that validate against a Pydantic schema, or the run fails loudly.
3. A reviewer with veto. The director agent rejects drafts with a written critique, and the drafter rewrites only the parts it criticised.

   A reviewer with veto. The director agent rejects drafts with a written critique, and the drafter rewrites only the parts it criticised.
4. Fast inference makes loops affordable. Running Llama-3.3-70B on Groq keeps a three-round review loop within seconds.

   Fast inference makes loops affordable. Running Llama-3.3-70B on Groq keeps a three-round review loop within seconds.

pages, report to briefpages, report to brief

100→2100→2

100→2100→2

specialised agentsspecialised agents

55

55

verified facts per brief, schema-checkedverified facts per brief, schema-checked

55

55

• Quality is judged by the reviewer agent and by hand; there is no labelled benchmark yet.

  Quality is judged by the reviewer agent and by hand; there is no labelled benchmark yet.
• Tables and charts in source PDFs are read as text and can lose structure.

  Tables and charts in source PDFs are read as text and can lose structure.
• Next: citation links from every claim back to the source page.

  Next: citation links from every claim back to the source page.