Skip to content
FI

Research

Some of our work is easier to understand as a question than a product. We keep the experiment, including the parts that tell us we were wrong.

RESEARCH — 01

VERDANT

A falsifiable recommendation engine for sparse, subjective preferences.

2026 · RECOMMENDER SYSTEMS · PAUSED AFTER v0.8 CONFIRMATION

v0.8 completed its preregistered confirmation experiment. Cardinal scoring improved over the historical formulation across all eleven simulated worlds, but the full policy failed six of eleven confirmation screens. The held-out AUDIT was deliberately not run.

READ THE RESEARCH →

VERDANT / 0.8

DEVELOPMENT_08
COMPLETE
SCREENS
05 / 11
VALIDITY FAILURES
00
AUDIT
WITHHELD
STATE
PAUSED