**Figure 13. The seven-step evaluation pipeline, and the human gate closing each step.**
Each column is one pipeline step (S1 to S7); each row is a role. **a**, Stage 1 is
deterministic: the three steps read a named source record, run a query with no model
in the path, and close at a human gate. The LLM row is empty throughout, which is the
figure's principal claim: no generative model produces any value reported from Stage
1. The corpus is frozen at G3 (bold), and that freeze is the stage boundary. **b**,
Stage 2 is LLM-assisted: each step names the model it uses and is closed by a gate
that checks the output back against the frozen Stage 1 corpus. Scoring at S6 applies
the standardised audit questionnaire (Q1 to Q7) and is author-applied; the model
summarises verified metrics only. A gate can refuse and return a step; it is not a
sign-off. The seven pipeline steps S1 to S7 and the seven audit questions Q1 to Q7
are distinct sets and are never numbered against each other. Drawn from
`deposit/pipeline.json`, the single source of truth for the step table, which carries
the SHA-256 of the deposited Supplementary Table 4.