# Session protocol for the four LLM judges (v9.2) This is the procedure each judge was run under. It replaces the tool-specific runbooks of the campaign, which contained only installation commands, folder names and launch snippets; nothing in them changed the task. The task itself is defined in the prompt files next to this document. ## Task Segment each of the 50 CLAUDETTE contracts into thematic clauses and assign **one theme per clause** from the closed vocabulary of the prompt (section 3 or 4, depending on the file). For every sentence the judge returns the theme, a block identifier and whether the sentence opens a new clause; at every clause opening it returns a short structured justification citing a literal fragment of the sentence. The legal nature of clauses was **not** asked of the judges: it was derived afterwards by a deterministic script from the theme, never by a model. ## Inputs a judge could see - the prompt file and this protocol; - the contract, one tokenised sentence per line, exactly as in the CLAUDETTE distribution (`.txt`; line *i* is sentence `id = i − 1`); - a deterministic map of the document (`_docfeatures.json`: number of sentences, indices of short title lines, an estimate of the number of blocks), computed by a script without any model; - its own output files. Nothing else: no annotation by a human or by another judge, no earlier session, no log of the campaign, no version history. Each judge ran in a fresh, isolated session with the outputs of the other judges moved out of reach before the session started. ## Procedure 1. **One document per run.** The judge reads the whole contract, writes a document plan (anticipated segments), then labels every sentence. 2. **Structural validation, blocking.** A validator checks every output file: exactly one entry per non-empty line of the contract, identifiers `0 … n−1` in order, theme in the vocabulary (or an `OTHER_