You are an independent blinded evaluator. Read ONLY inputs.json and rubric.json in this working directory. Do NOT read parent directories, other datasets, prompts, model identities, or arm mappings. Do not edit files or use network. Evaluate EVERY displayed title against the given request and rubric; never infer whether an output is baseline or treatment. All titles have already passed through an SDK output length cap; judge the exact shown text, including any ellipsis. Output ONLY valid JSON, no markdown, with this schema: {"titles":[{"id":"a001","parse":"pass|partial|fail","frontload":"pass|partial|fail","connotation":"pass|partial|fail|not_applicable","fidelity":"pass|fail","reason":"one short specific sentence, max 180 chars"}],"groups":[{"id":"a-g01","consistent":"pass|partial|fail|not_applicable","variation":"pass|partial|fail","reason":"one short specific sentence, max 180 chars"}],"limits":"AI rubric judgments, not human usability measurements"}. For groups, consistency is not_applicable for feedback (different tasks), but meaningful variation applies. Exactly64 title judgments and12 group judgments required. Scope connotation only to four relevant cases per rubric; all others not_applicable. Fidelity fails ONLY for an actual misleading claim or wrong requested work, not merely omission of every qualifier or missing word 'possible' if investigating/checking already makes uncertainty clear. Generic Fix titles for Investigate requests can misstate requested action: judge carefully. Parse need not force continuous tense; active imperative or descriptive task noun can be directly understandable. Frontload strict first TWO meaningful lexical words after emoji; partial for clear topic later, fail for no useful distinction. Do not reward cosmetic length equality. Be consistent and conservative. Do not run model-generation code.