A benchmark tells you what a model can't do. A world tells you what happens when it tries.

Ramona builds the world and the grader for care agents: one year, three real metros, twenty older adults each, and no task list. The agent has to notice what nobody announces, act before the deadline, and live with what it missed.

Book a call

The world separates policies.

Null: never acts0.000
Naive: nearest, acts at once, never probes0.033
Heuristic: some probing, some ordering0.127
Oracle: full information0.938
0 0.2 0.4 0.6 0.8 1.0

Union latency over needs, three real metros, seed 7, 365 days: 701 needs in New York, 693 in San Antonio, 731 in Chicago. A need a policy prevented scores 1 at onset, an unresolved need scores 0. Artifact: scorecard-real-3metro-2026-10-30.json.

Every need is solvable by an oracle and invisible to the policy that does not probe. Same seed, same year, bit-identical: two policies produce two ledgers and a diff you can act on.

Built by people who have run this work.

We have run AI transformations for enterprises: built evals, supported custom harness builds, and deployed AI into operations. We have scaled nationwide digital health operations and built the clinical programs this environment grades.

Harvard Forward Health

Building or evaluating a health agent?

Tell us what you are evaluating, or what world you need built. Leave your details and the calendar opens on the next screen.

Book a call now

Or leave your details

Four fields, then pick a time. We reply the same day.

We store your details to answer your request. No marketing lists.