Does Tessary truncate a long trace before grading it?

Partly: two classifiers read a bounded window, while triage and RCA, the LLM steps behind them, include fewer items rather than clip any one of them.

In the LLM lanes the bound is by count, not by content. A tool result or a model completion either goes in whole or doesn’t go in at all. Grading a clipped payload is worse than grading fewer items, because the one line that was wrong is usually the line that got cut, and the verdict comes back confident anyway.

The two classifiers that read text through a model clip on purpose. Frustration reads one user message and the four before it, swaps pasted blocks for a marker, and cuts a long message to its head and tail. Groundedness reads an answer and its retrieved documents together in one pass of up to 8,192 tokens, per its model card. A very long answer is cut at a fixed length, the documents are shortened to fit beside it, and an answer that still leaves no room for them isn’t scored at all.