# PRAXIS analytical-quality evaluation protocol

Protocol version 1.0, September 30, 2026. Proposed study, not published results.

Use six to ten public or permissioned policy/narrative tasks. Freeze and hash the same source set for every condition. Compare the team's current workflow, a named general-purpose model with the same sources, and PRAXIS with reviewed context. Record model/version, prompt, retrieval configuration, tool use, date, latency and cost. Avoid assigning a PRAXIS advantage before running the comparison.

Use blind reviewer scoring where feasible. Independently inspect original source passages. Score unsupported claims, citation entailment and source attribution, temporal interpretation, commitment versus implementation, preservation of disagreement, useful qualifications, and time to a reviewed brief. Record corrections and unresolved evidence; do not treat provider grounding segments as verbatim source extracts.

Predefine task-level acceptance criteria. Publish per-task results, failures and sample limitations, along with sanitized inputs, scoring instructions and rerun commands. Small synthetic examples establish neither representative performance nor institutional suitability. Obtain permission before sharing customer material. Report both reviewer agreement and disagreements; a structural validation pass is a separate result.