New Context: PolicyEngine is also implementing CiviCRM for constituent relationship management. This significantly changes the landscape for non-technical work (grants, papers, partnerships).
Key Insight: CiviCRM has built-in grant tracking, potentially eliminating the need for separate grant management tools!
GitHub Projects v2 uses GraphQL, but filtering by custom fields (like due dates) in the API is extremely cumbersome:
# GitHub Projects v2 - Complex and Limited
For PolicyEngine, Linear is the better choice due to:
| Fiscal year | Lagged RPI used | Baseline | Policy | HMRC revenue impact (£m) |
|---|---|---|---|---|
| 2024–25 | — (frozen) | £85,000 | £90,000 | –150 |
| 2025–26 | — (frozen) | £85,000 | £90,000 | –185 |
| 2026–27 | 2024–25: +2.0% | £87,000 | £90,000 | –125 |
| 2027–28 | 2025–26: +2.5% | £89,000 | £90,000 | –50 |
| 2028–29 | 2026–27: +3.0% | £92,000 | £90,000 | +65 |
| ╭───────────────────────────────────────────────────────╮ | |
| │ ✻ Welcome to Claude Code! │ | |
| │ │ | |
| │ /help for help, /status for your current setup │ | |
| │ │ | |
| │ cwd: /Users/maxghenis/PolicyEngine/policyengine-us │ | |
| ╰───────────────────────────────────────────────────────╯ | |
| > /encode-policy is running… | |
| ⎿ Interrupted by user |
December 2024
Government benefit calculations affect millions of lives. A miscalculated food assistance benefit isn't just a bug—it's a family's grocery budget. An incorrect tax credit isn't just an error—it's someone's ability to pay rent. That's why at PolicyEngine, we've developed a revolutionary approach to ensure AI-generated code accurately implements government regulations: isolated multi-agent development.
When you ask an AI to implement a government program, it faces a fundamental challenge. Given a test that expects a benefit of $500, the AI might: