{
  "measurement": "citation-triage-agent-and-human",
  "date": "2026-08-03",
  "question": "Given a reading list, how much can an AI agent turn into citable records on its own, and what does it have to hand back?",
  "method": {
    "endpoint": "https://provinglab.dev/mcp",
    "tool": "extract_citation",
    "sources": 12,
    "selection": "A mixed list as it would appear in a term paper: publishers behind bot defences, open access, a repository, a preprint server, EU law, an encyclopaedia and a bare DOI. Not chosen for whether they work.",
    "control": "Every refusal was re-fetched from an unrelated network to confirm the block is real."
  },
  "results": {
    "handled_by_the_agent": 8,
    "handed_back_to_the_human": 4,
    "seconds_total": 13.1,
    "seconds_per_source": 1.09,
    "records_complete_with_ris": 8,
    "handed_back": [
      {
        "host": "www.mdpi.com",
        "reason": "http-403"
      },
      {
        "host": "www.sciencedirect.com",
        "reason": "http-403"
      },
      {
        "host": "www.ssoar.info",
        "reason": "bot wall served to server-side readers"
      },
      {
        "host": "eur-lex.europa.eu",
        "reason": "near-empty response to server-side readers"
      }
    ]
  },
  "limits": [
    "Read server-side without JavaScript; anything a page adds after loading is invisible.",
    "The split is not stable over time — publishers tighten bot defences, and EUR-Lex began returning an empty response to server-side readers between two runs on the same day.",
    "Two of the eight records carry no author because the page declares none. That is the page's gap, not the reader's, and it is reported rather than filled in."
  ],
  "license": "CC-BY-4.0"
}