Scores each article twice — once under each candidate analysis prompt — and reports the difference. Nothing is persisted: no cache write, no facts-KB write, no outlet stats, so a run cannot contaminate production data. Research runs once per article and both variants receive the identical context, and prior-facts KB context is omitted, so the only thing that differs between A and B is the prompt. Single model, no ensemble. Only the deterministic grounding layers apply (L1 score recompute, L2 research-contradiction penalty).
Single-model scoring carries roughly ±14 points of run-to-run noise — use several articles and 2–3 repeats before trusting a delta.