Methodology

Each virtual candidate is authored from a candidate-specific kernel document. The machine-readable candidate fixture is derived from that kernel, not hand-maintained separately.

Each debate artifact follows a fixed verification-first path: golden spec, frozen fixtures, implementation, replayable verification, and evidence. Session outputs are written as JSONL, rendered to static HTML, and accompanied by runtime metadata, critique metadata, response-contract checks, and evidence records.

During generation, the model directly receives system constraints, candidate kernel stance, topic context, prior-round carry-forward, and response rubric. It does not directly read the full public site, the full source-reference prose, or unrelated project documents during a normal run.

Source references support the authored kernel, but they are not direct live retrieval inputs in V01.

V01 currently covers seven virtual Texas Senate candidates, three-round prototype sessions, explicit critique metadata, and guardrail checks. It is designed for inspectability before scale.