Adjudication: why one model isn't enough
Sending every request to an LLM-judge is slow, expensive, and non-deterministic. A layered pipeline resolves most requests without a model at all.
The naïve guardrail is a single model call: ask an LLM whether a request is safe, then act on the answer. It works in a demo and falls apart in production. It adds hundreds of milliseconds to every request, costs a model call per evaluation, and gives you a different answer on a Tuesday than it did on a Monday.
Adjudication is a pipeline, not a call. Deterministic Edict rules resolve the clear cases first - a denied tool, an SSN in an output - in microseconds, with zero model cost and a perfectly explainable reason. A classifier ensemble scores what the rules can't decide cleanly. Only genuinely ambiguous requests escalate to a model-graded review.
The payoff is that the overwhelming majority of traffic never touches a model on the enforcement path. You get the latency and determinism of rules where rules suffice, and the judgement of a model exactly where you need it - and nowhere else.
This is the honest version of 'multi-agent validation': not a swarm of models second-guessing each other, but a disciplined escalation that spends compute only on uncertainty.
Further reading
