consensus

Apply self-consistency to a structured prompt: run it N times, parse each output as JSON, fuzzy-match findings across passes, keep only what converges.

The vote is deterministic — a Python script consolidates the JSON outputs without any LLM judge. The schema and vote key are agreed upfront and cached, so repeat invocations on the same prompt skip the schema-design step.

Arguments

• <prompt> (positional, required): the structured prompt. Must describe a task whose output can be a JSON list/object (findings, extractions, classifications). Free-form text generation is not supported — the vote needs structure.
• --passes N (optional, default 3): how many times to run the prompt. Must be ≥2.
• --threshold M (optional, default ceil(N/2)): minimum votes for a finding to be retained. Must satisfy 1 ≤ M ≤ N.
• --force-fresh (optional flag): bypass the schema cache and re-derive the schema from scratch.

Step 0: Validate inputs

PASSES=${PASSES:-3}
THRESHOLD=${THRESHOLD:-$(( (PASSES + 1) / 2 ))}