Lancet

Lancet is an evaluation system for agents and models. It follows a familiar discipline — the way pytest made tests a first-class part of writing software, Lancet makes evaluation a first-class part of shipping an agent. It consumes the traces other tools already produce and turns them into measured, defensible judgements about behaviour.

Generative systems demo well and fail silently. An agent can appear to work and be wrong in ways no one sees, because the failure hides inside a fluent, confident output. Without evaluation grounded in evidence, teams ship on hope.

Arc's position is that the unsolved problem of the GenAI era is not capability but trust — and trust is earned by measurement, not assertion. Lancet is the instrument that measures: evidence-grounded evaluation and calibration, with the core method protected. It is where Arc's discipline of telling the truth about a system becomes a tool.

Alpha