Verifier selects, gates decide

The rule. An LLM verifier slots above the deterministic gate, never in place of it. Its province is graded selection where no deterministic check exists — ranking candidates, best-of-N picking, monitoring a long run's progress. Admission to the record is decided by deterministic gates (tests, preflight, lint); a probabilistic verdict is a selector and an early-warning signal, not an authority.

Why. Graded selection is where LLM judgment measurably helps, and gate replacement is where it measurably fails. The LLM-as-a-Verifier study (Kwok et al., Stanford/Berkeley/NVIDIA, 2026) ranks agent trajectories with an LLM but keeps a hidden deterministic grader as ground truth throughout: verifier pairwise accuracy tops out near 78%, so a gate built on it alone silently passes roughly one wrong candidate in five, while oracle best-of-K reaches near 99% — the selector, not the generator, is where the judgment pays. The same study's graded score over a run's prefix separates healthy from stalled runs, explicitly as advisory telemetry rather than a kill switch. The GLM-5 report (Zhipu AI and Tsinghua, 2026) draws the complementary line: its agentic judge counts only after measuring 94% agreement against human labels, and when a learned reward was gamed the fix went to the measuring instrument's grounding, never the outputs — a gate handed to a probabilistic judge inherits gaming surface and calibration drift as standing liabilities.

Bounds.

• An LLM verdict may advise a gate — flag, rank, annotate, propose — but the gate's blocking decision stays deterministic. Where no deterministic check exists yet, the verifier is the interim and building the check is the debt.
• Claims backed only by LLM judgment are marked judged-only and never reported in the vocabulary of enforcement (composes with enforcement-claims-are-facts).
• Discrete verdicts are for gating a single artefact; ranking uses graded scores — the study's discrete judge tied on 88 of 100 comparisons where a graded expectation ranked correctly.

Live instance. The 2026-07-08 multi-agent record review used independent LLM refuters to grade raw findings before capture — 3 of 46 refuted and dropped, the rest clustered into ledger issues — while admission of the resulting changes stayed with make preflight and the record's deterministic gates: selection above, gates deciding.

Promotion. The MVP is a convention marking judged-only verdicts as such wherever they persist (ledger entries, review notes), plus a labelled fixture set of known-good and known-bad artefacts attached to any oracle whose scores feed selection. The tool is an opt-in oracle adapter tier whose measured accuracy and tie rate are recorded per configuration — demotable to advisory on stale calibration, archivable when a deterministic check supersedes it.