Raycaster/ Eval

Company worlds

Step inside the desks our agents work.

Step inside the company desks our benchmarks are built on — the documents, assignments, and scored agent work, published in full.

A world is a complete company environment — file rooms, controlled registers, and colleague chat — rebuilt so every document, assignment, and scored agent run can be published for inspection.

  1. Broad Institute — Divalent siRNA IND ResponseBuilt from a real public FDA recordA real partial clinical hold, rebuilt verbatim from the Broad Institute’s public IND filing — 294 real files. Frontier models did the work; the best still authorized activity FDA had frozen.294 real filesPublic FDA recordScored model runs
  2. Havenor TherapeuticsA fictional company we can publish in fullA simulated pharmaceutical desk — 853 files, working systems, and 120 colleagues on chat who answer in character. Models sail through the routine and fail the judgment calls.853 files120 in-character colleaguesScored model runs