)]}'
{
  "commit": "ffe28aa77ec85eeabb10aaedda5e4c712ea8eb44",
  "tree": "37df71256b2474d7298c24ff6f2706e570ac53f6",
  "parents": [
    "c5143241fb90939fd0441cf3dcf6313946a4cdd7"
  ],
  "author": {
    "name": "Marc Kupietz",
    "email": "kupietz@ids-mannheim.de",
    "time": "Sat Sep 05 17:50:45 2026 +0200"
  },
  "committer": {
    "name": "Marc Kupietz",
    "email": "kupietz@ids-mannheim.de",
    "time": "Sat Sep 05 19:41:49 2026 +0200"
  },
  "message": "DocPrompting: Prompt the way a user would\n\nWhat these tests send is also what we implicitly recommend, and it had\ngrown into something nobody would type: the task appeared twice, once\nbefore the Readme and once after it, in two slightly different wordings,\nwith the documentation wedged in between. It now appears once, after the\ndocumentation, where an instruction belongs when the context is long.\n\nThe persona goes with it. \"You are an expert R programmer\" moves the\nwording of an answer rather than its correctness, and what was actually\nwanted - clear tidyverse code, in the style of the Readme\u0027s own examples\n- is now asked for. Asking for a single code block addresses at its\nsource what extract_r_code() has to repair afterwards.\n\nRemoving the repetition cost the collocation task its second mention of\n\"lemma\", which turned out to be carrying it: the answers fell back to the\nword form. That is a gap in the Readme rather than in the prompt, and is\nfixed in the commit before this one; three runs pass with both changes in\nplace.\n\nCo-Authored-By: Claude Opus 5 \u003cnoreply@anthropic.com\u003e\nChange-Id: I85c8256d6c066768bd816af28fd4c1c78f969169\n",
  "tree_diff": [
    {
      "type": "modify",
      "old_id": "f547c042a532b937901fda68b489e1e8c03c2a2f",
      "old_mode": 33188,
      "old_path": "tests/testthat/test-readme-against-llm.R",
      "new_id": "e37475fa130007afc1259418916806b054a70032",
      "new_mode": 33188,
      "new_path": "tests/testthat/test-readme-against-llm.R"
    }
  ]
}