Setup
One real decision: should a deep-tech fund invest in a ramjet-propulsion startup? Three runs of the same frontier model:
- Alone — unrestricted agentic run: internet, tools, sub-agents, no budget cap.
- With a research memo — the same run, plus a Deep-Research-style memo in context.
- With our data — the same run, plus ~1,000 of our generated decision sequences in context. Nothing trained; context only.
Each run's top 300 considerations were pooled, clustered, and blind-scored for decision impact on a 0–100 scale.
Result
The run with our data produced 92 exclusive high-impact considerations (impact ≥ 70) against 42 for the model alone and 33 with the memo — at higher average impact (73 vs 61) — and it alone reached the study's single top-scored insight (94/100). The memo added nothing over the bare run: the bottleneck is structure, not effort.
Why it matters
This is expertise lent temporarily, through context — the weakest possible way to deliver it. If holding a sliver of the generated space in context already shifts what a frontier model considers on a real decision, the natural question is what training on the full space does permanently. That is exactly the question the transfer work begins to answer.
This establishes
- Generated experience in context materially changes what a frontier model considers on a real decision — under blind scoring.
- The effect is not explained by added effort or added research: the memo arm controlled for that.
pooling, clustering and scoring protocol available on request · hello@truna.ai