Multi-Model response comparator

Coverage is judged by matching the words of each key point against the answer text, weighted so a partial word match counts half. That is deliberately shallow: a synonym the answer never states scores as missed, and none of it is a model's judgment. Reading ease is the Flesch formula over an estimated syllable count. Claims are sentences with absolute wording, causal links, or numbers; citations are URLs, markdown links, arXiv and DOI strings. Latency and cost appear exactly as you typed them, because nothing here calls a model or a clock.