A biological protocol can run for weeks: cell lines to keep, assays to stage, a plan that changes as results come in, and rules on sterility, dosage and sequence at every step. A model that is right 97% of the time is not good enough when the other 3% is a contaminated cell line.
This synthesis of our June research argues that frontier models have capability to spare. What is scarce is fidelity, meaning actions follow the rules a model knows, and durability, meaning that fidelity lasts across millions of turns. Grant compliance, underwriting and legal research have the same shape.
What we found
- Fidelity belongs to the whole system. Improving one part at a time leaves the gap open.
- Three failures the literature reports separately turn out to be one problem: drift when the context is compacted, the gap between what a model believes and what it does, and the wall capable models hit.
- On a regulated task, fidelity rose from 0.43 for a plain agent to 1.00 with Nenya.
- With Nenya, a regulatory rule broken once stays learned. A bare model makes the same breach every time the situation comes back.
- In the paper’s model, the horizon over which an agent stays reliable grows exponentially with compute.