public-records-gym-demo

A tool-use RL environment over joined Minnesota public records, by Brick and Mortar AI LLC (brickandmortar.dev). An agent answers questions about land, buildings and businesses by querying seven public-record layers through an ArcGIS-style API (2,000 rows per call, no GROUP BY, no SUM). Every answer is computed from a frozen snapshot by SQL, re-derived through the agent's own tools, and graded without a model.

The tasks are built around the ways the obvious query is wrong, and the traps are true in the rows and written in no layer note: a fee repeated on every parcel row of one permit, two assessed-value columns that disagree on 333 sales, county-scoped parcel ids that collide across Carver and Scott, permits completed before they were issued.

This is the public demo: 2 tasks over only the permit rows they read. The full environment (227 tasks across 7 templates and 7 layers, 1.97M rows, a held-out test split with a measured frontier baseline) is licensed commercially: aidan@brickandmortar.dev.

prime env install <owner>/public-records-gym-demo
uv run vf-eval public-records-gym-demo -n 2