Research · Main Labs
Running large language models across a mix of tiers promises lower cost and latency by resolving easy queries locally and escalating hard ones upstream. That escalation hinges on a confidence signal the small model emits about its own output. This paper shows the binding constraint inverts: where compute is billed, cost decides; where it is sunk on owned hardware, signal quality decides.
Agentic workloads run for a long time but sit idle in tool calls for most of it, which makes reserving capacity for them expensive. This paper puts that reservation on a pre-registered test bench, measuring across 108 configurations against six equally tuned baselines whether per-flow completion guarantees buy more completed work per provisioned GPU-hour than they strand. Throughput alone does not carry the reservation, but the guarantees do deliver interactive tail-latency isolation and bounded containment of runaway flows.
Agent tasks are increasingly delegated across organizational boundaries, and existing work identifies the absence of a way to express capability policy between domains as a gap. This document defines four policy attributes for inter-domain agent delegation, the declarations each attribute carries, and a validity condition on delegation chains that no single party establishes by observing the whole chain.
Controls that live inside an AI system's own runtime, such as system prompts, output filters and guardrail libraries, can be reached by anything that influences that runtime, including the system itself. This paper argues that such escapable systems need alignment enforced at execution time, outside their trust boundary, and states five properties that layer must satisfy, adding revocable compute to four authorization properties. It presents an open-source Rust kernel that realizes all five, and evaluates it in a pre-registered campaign in which the governed system's own self-improvement loop is tasked with defeating it.