You use an LLM to judge your LLM outputs. That is circular — the judge shares the failure modes of the judged. How is that not theater?

The full executive answer

It would be theater if the judge were the foundation, so we built it as the roof instead. The load-bearing layers are deterministic and share no failure modes with any model: schema validation on every output, content rules that check citations against a real catalogue and numbers that must reconcile, numeric provenance tracing every figure to input, guardrail consistency checks, and a mutation-tested regression harness — code, not opinion. A judge hallucination cannot corrupt any of those gates, because none of them consult a model.

The judge sits on top for what deterministic rules genuinely cannot measure — is the analysis specific rather than generic, is the register boardroom-grade, is thin evidence hedged honestly — and it is constrained against the known LLM-judge pathologies: a fixed six-dimension rubric with defined one-to-five anchors rather than free-form scoring; a forced quotation of the weakest sentence in the output, which makes lazy blanket scores detectable; comparison against a committed baseline so what fails a build is relative regression, not an absolute vibe; and structural separation between judged output and judging pass. I will freely grant the residual limitations — self-preference bias and rubric gaming are real literature — which is why judge scores gate quality trends rather than truth claims.

And the escape from circularity has a designed endpoint: human signal. Down-rated, edited and flagged outputs from real users are the explicit candidate pool for human-anchored gold cases, and the calibration loop grades forecasts against measured reality — no model in that loop at all. Deterministic floor, constrained judge in the middle, human-and-reality anchoring accruing on top. That is the honest architecture for evaluating generative systems in 2026; anyone claiming better is skipping a layer somewhere.