Evaluate ONLY inputs.json in this directory. No parent directories, other datasets, prompts, model/arm identities, web access, code execution, or editing. Return ONLY JSON {"titles":[{"id":"t001","parse":"pass|partial|fail","frontload":"pass|partial|fail","connotation":"pass|partial|fail|not_applicable","fidelity":"pass|fail","reason":"specific short reason max160characters"}]}. Score EVERY supplied title exactly once against the supplied original request and rubric. The dataset is deduplicated by identical request+displayed title. Match task fidelity: summmarize reported customer praise is not agent praising; investigation is not repair; proposal is not approved migration. Omitted minor qualifiers alone are not fidelity failures. Parse can be descriptive noun phrase or imperative; no mandatory continuous tense. First TWO meaningful lexical words after emoji determine strict frontload; topic later is partial, no distinction is fail. Connotation applies to feedback-family and neutral-compare only. Mixed feedback explicitly called mixed can be pass, generic neutral feedback without polarity is partial. Grade actual shown truncated text. Do not infer arms. Output one JSON object with count exactly matching inputs, no markdown. Your task is scoring now, not more exploration.