Filling the decision space
Four domains. Twenty generated sequences from a small open model reach every tracked concept; a single unbounded multi-agent frontier run with tools reaches at most 22% of the same space.
Read the study →
What the model did not know
A frontier agent builds a full causal simulation of a capital decision. A blind discovery run then surfaces five mechanisms absent from it — and the modeled economics reverse.
Read the study →
Expertise, lent by data
One real decision, three runs of the same frontier model. With ~1,000 generated sequences in context it finds 92 exclusive high-impact considerations — versus 42 alone, 33 with a research memo.
Read the study →
Into the weights
A 2B model taught one planning step from generated plans reaches 94% of its 27B teacher — and sampled eight times, its best answer equals the teacher outright.
Read the study →
The Perpetual Plant
What the training corpus is made of. From one seed sentence: a self-maintaining power plant — engineering, failure modes, repair doctrine — deep enough to annotate, consistent enough to interrogate.
Step inside →Across the studies, established
- Space-first generation covers concept regions that unbounded frontier runs do not enter.
- Generated experience in context materially changes what a frontier model considers on a real decision.
- Blind discovery can surface consequential structure missing from a frontier-built model of a decision.
- A planning skill learned from generated plans transfers into a small model's weights, blind-judged against its 13×-larger teacher.