← Research

What the model did not know

Monte Carlo answers the uncertainty you specify. It cannot, by itself, discover a mechanism that was never represented. This study isolates that second problem — model-space uncertainty — and tests whether blind discovery can close it.

illustrative scenario blind protocol simulation-gated

Setup

An unrestricted agentic frontier run — browsing, tools, sub-agents, no preset budget — was given a staged, multi-billion-dollar EV manufacturing programme (illustrative: every numeric input assumed, no company data used) and built the baseline causal simulation itself: 36 variables, 20,000 futures over 40 quarters.

Our discovery system then ran blind: it received the decision problem, but not the baseline's variables, causal graph, or omissions. Its findings became candidate variables only if re-simulation showed they materially changed the decision — impact gated by arithmetic, not by one model rating another's prose.

Result

The blind search produced 22 findings; seven became candidates; five survived the simulation gate — none of them exotic: a total-cost-of-ownership demand cliff, step-function grid reinforcement costs, a charging-anxiety premium, grid interconnection delay, and a resale-liquidity penalty. Persistent regimes and structural frictions — the kind that change the central case, not just the tail.

Incorporated into the same simulation machinery, they reversed the modeled economics of the decision: the median outcome flipped sign, and the value of preserving the option to halt rose several-fold. The result is not five more risk bullets; it is a different model of the world.

Two constellations of a known model, and mechanisms beyond its boundary
The shape of the finding: a coherent model of the decision — and consequential mechanisms sitting beyond its boundary, unreachable from inside it.

Why it matters

Frontier models are strong at reasoning inside a supplied representation. The harder question is whether they reliably construct the right representation. A polished answer can be confidently wrong when the relevant mechanism never entered the model — and the practical conclusion changes from "forecast better" to "structure the commitment so you can change your mind."

This establishes

  • A blind discovery process surfaced mechanisms absent from a frontier-model-built baseline.
  • Those mechanisms translate into executable causal structure and materially change the decision distribution.
  • Discoveries were gated by simulated impact, not by model-graded prose.

full study — mechanisms, distributions, audit trail — available on request · hello@truna.ai