Executive summary
A biological protocol can take weeks: cell lines to maintain, assays to stage, a plan that changes as results come in, with sterility, dosage and sequence rules binding 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 brings Nenya’s June research into one account of why agents miss that bar. The frontier has capability to spare. Two properties are scarce: fidelity, meaning actions follow the rules a model knows, and durability, meaning fidelity lasts across millions of turns. Grant compliance, underwriting and legal research share the same shape.
Findings
- Fidelity is a property of the whole system. Improving one part at a time leaves the gap open.
- Three failures the literature reports separately (drift when context is compacted, the gap between what a model believes and what it does, and the wall capable models hit) turn out to be one problem.
- On a regulated task, fidelity climbs 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 recurs.
- In the paper’s model, the reliable horizon grows exponentially with compute.
Those are the properties Nenya adds to the agents and models you use, for as long as the work runs.