Ordinary generation circles what it already knows. The question this pilot asks is whether a machine that maps the space as it generates can reach the regions a strong agentic run never enters.
Four domains, each seeded with a single question. Our machine generated 20 decision sequences per domain from a small open model (Qwen-class 27B, 4-bit). The baseline: a single unbounded multi-agent frontier run with tools — a frontier model in a multi-turn agentic harness, free to browse, recurse and spawn sub-agents (169 samples). Both outputs were mapped against a tracked concept space per domain, and coverage was measured, not eyeballed.
The generated sequences reached every tracked concept in all four domains. The frontier run reached at most 22% of the same space — staying, domain after domain, near where it started. At pilot cut-off our coverage curve was still climbing: the probe set was full; the space was not.
Coverage is the property everything else depends on. A specialist trained on data that circles its seed inherits the seed's blind spots; a specialist trained on data with measured coverage of the mapped space can hold experience for regions where nothing has happened yet. This pilot is the first direct measurement that our generation reaches those regions and a strong agentic baseline does not.
full coverage maps, probe sets and run logs available on request · hello@truna.ai