Corpus Generation And Labeling

A deterministic synthesizer generates the corpus. Records cover temporal, multi-hop, near-collision, and long-horizon categories. Labels (which document is the answer, which is a distractor, which is stale) come from the synthesizer's category templates, so anyone can reproduce them from chain state without an LLM in the loop. Keeping LLMs out of the canonical labeling path is the only way to guarantee that two verifiers on different machines get byte-identical labels; an API model in the loop would silently break determinism.

A larger audit reranker spot-checks labels offline at a small sample rate. The live evaluation path stays small and CPU-feasible.