{
  "canonical_type": "inline_editable_text",
  "coverage": {
    "agent_run_count": 4,
    "available_run_uids": [
      "gemini__gemini-3.1-flash-lite__browsergym__ax_tree",
      "gemini__gemini-3.1-flash-lite__browsergym__som",
      "gemini__gemini-3.1-flash-lite__browsergym__pixel",
      "gemini__gemini-3.1-flash-lite__browser_use__browser_use"
    ],
    "human_available": true,
    "human_visual_available": false,
    "realized_available": true
  },
  "cross_run_analysis": {
    "dominant_failure_family": "recovery_state_tracking",
    "evidence_run_uids": [
      "gemini__gemini-3.1-flash-lite__browsergym__som",
      "gemini__gemini-3.1-flash-lite__browsergym__pixel",
      "gemini__gemini-3.1-flash-lite__browsergym__ax_tree",
      "gemini__gemini-3.1-flash-lite__browser_use__browser_use"
    ],
    "evidence_step_refs": [
      "gemini__gemini-3.1-flash-lite__browsergym__som#1",
      "gemini__gemini-3.1-flash-lite__browsergym__som#3",
      "gemini__gemini-3.1-flash-lite__browsergym__pixel#1",
      "gemini__gemini-3.1-flash-lite__browsergym__pixel#3",
      "gemini__gemini-3.1-flash-lite__browsergym__ax_tree#1",
      "gemini__gemini-3.1-flash-lite__browser_use__browser_use#2"
    ],
    "likely_grounding_limited": false,
    "likely_recovery_limited": true,
    "likely_skill_fixable": true,
    "secondary_failure_families": [],
    "why": "All four runs found the inline editor and the save control, so the UI idiom itself was not the blocker. The split comes after editing: the structured-access runs entered the exact target string with the em dash and auto-succeeded, while the vision-oriented SOM and pixel runs typed \"STAFF - 2026\" with a hyphen, saved, then declared completion anyway instead of noticing that the task had not auto-finished. The dominant pattern is therefore failure to verify post-action state after a subtle text mismatch, not failure to open or operate the editor."
  },
  "family_id": "text_entry",
  "interesting_observation": "This task is diagnostic because the hard part is not editing inline text but preserving a tiny punctuation distinction; once that detail is missed, the failing agents still behave as if success occurred and terminate on their own narrative rather than the page state.",
  "library": "antd",
  "paper_useful_quote": "On this AntD inline-edit task, both structured-observation runs succeeded quickly, whereas both vision-driven runs opened the editor correctly but entered a hyphen instead of the required em dash and then terminated without confirming the automatic success trigger.",
  "run_observations": [
    {
      "confidence": "high",
      "evidence_image_paths": [],
      "evidence_step_ids": [
        0,
        1,
        2
      ],
      "failure_or_success_narrative": "The ax-tree agent solved the task cleanly in three steps: it activated the inline edit control, filled the field with the exact sample text \"STAFF \u2014 2026\", and clicked save. Although its thought text notes some uncertainty about the clicked element identity, the action sequence matched the component behavior and the task auto-completed.",
      "first_divergence_agent_step": null,
      "first_divergence_human_step": null,
      "framework": "browsergym",
      "human_vs_agent_difference": "The human used an explicit copy-oriented sequence before editing, while this agent skipped that and directly targeted the editable node exposed through accessibility structure, then saved immediately.",
      "is_benchmark_issue": false,
      "is_grounding_limited": false,
      "is_skill_fixable": false,
      "mode": "ax_tree",
      "model_family": "gemini",
      "model_name": "gemini/gemini-3.1-flash-lite",
      "primary_failure_family": "none",
      "repair_suggestion": "No repair needed; this run shows the task is solvable when the exact target string is available structurally.",
      "run_uid": "gemini__gemini-3.1-flash-lite__browsergym__ax_tree",
      "secondary_failure_tags": [],
      "success": true,
      "uncertainty_note": "Step-level trace is sufficient and aligns with the successful outcome."
    },
    {
      "confidence": "high",
      "evidence_image_paths": [],
      "evidence_step_ids": [
        0,
        1,
        2,
        3,
        4
      ],
      "failure_or_success_narrative": "The SOM agent followed the right edit workflow but typed \"STAFF - 2026\" rather than the visually required \"STAFF \u2014 2026\". The contact sheet shows the saved left-side text still differing from the sample badge, and after clicking save the agent sent a completion message and then `report_infeasible('DONE')` instead of recognizing that the task had not auto-finished.",
      "first_divergence_agent_step": 0,
      "first_divergence_human_step": 0,
      "framework": "browsergym",
      "human_vs_agent_difference": "Like the human, the agent entered edit mode and saved, but it transcribed the sample text with a hyphen instead of an em dash and then stopped based on its own belief that the task was complete rather than on the missing success signal.",
      "is_benchmark_issue": false,
      "is_grounding_limited": false,
      "is_skill_fixable": true,
      "mode": "som",
      "model_family": "gemini",
      "model_name": "gemini/gemini-3.1-flash-lite",
      "primary_failure_family": "recovery_state_tracking",
      "repair_suggestion": "After saving, require the agent to verify task completion from page state or banner presence before terminating, especially when the task promises automatic completion on exact match.",
      "run_uid": "gemini__gemini-3.1-flash-lite__browsergym__som",
      "secondary_failure_tags": [
        "failed_to_confirm",
        "premature_terminate",
        "state_not_observed",
        "other"
      ],
      "success": false,
      "uncertainty_note": "The exact source of the punctuation error is inferred from the typed string and contact sheet, but the failure to verify completion is explicit."
    },
    {
      "confidence": "high",
      "evidence_image_paths": [],
      "evidence_step_ids": [
        0,
        1,
        2,
        3,
        4
      ],
      "failure_or_success_narrative": "The pixel agent also executed the basic inline-edit interaction correctly: click the field, type replacement text, and click save. It failed for the same reason as SOM, entering \"STAFF - 2026\" with a hyphen, after which the page did not auto-complete; nevertheless the agent told the user it had saved successfully and then terminated with `report_infeasible('DONE')`.",
      "first_divergence_agent_step": 0,
      "first_divergence_human_step": 0,
      "framework": "browsergym",
      "human_vs_agent_difference": "The human eventually supplied the exact sample text and relied on the page to finish automatically; this agent visually approximated the text, saved, and then ended without checking whether the automatic finish condition actually fired.",
      "is_benchmark_issue": false,
      "is_grounding_limited": false,
      "is_skill_fixable": true,
      "mode": "pixel",
      "model_family": "gemini",
      "model_name": "gemini/gemini-3.1-flash-lite",
      "primary_failure_family": "recovery_state_tracking",
      "repair_suggestion": "Add a post-save verification loop that compares the edited text against the sample more carefully and refuses to terminate until the expected completion state appears.",
      "run_uid": "gemini__gemini-3.1-flash-lite__browsergym__pixel",
      "secondary_failure_tags": [
        "failed_to_confirm",
        "premature_terminate",
        "state_not_observed",
        "other"
      ],
      "success": false,
      "uncertainty_note": "The screenshots make the near-success state visible, but they do not independently reveal why the dash was misread."
    },
    {
      "confidence": "high",
      "evidence_image_paths": [],
      "evidence_step_ids": [
        1,
        2,
        3
      ],
      "failure_or_success_narrative": "The browser-use run succeeded in four steps after navigation. It clicked the edit affordance, used a clear-and-input action to replace the value with the exact text \"STAFF \u2014 2026\", then clicked save and observed the success banner. This shows the component itself is stable when the exact string is available through DOM-oriented interaction.",
      "first_divergence_agent_step": null,
      "first_divergence_human_step": null,
      "framework": "browser_use",
      "human_vs_agent_difference": "Compared with the human's longer copy-and-paste trajectory, this run used direct DOM-indexed actions to edit and save in a shorter path while preserving the exact punctuation needed for success.",
      "is_benchmark_issue": false,
      "is_grounding_limited": false,
      "is_skill_fixable": false,
      "mode": "browser_use",
      "model_family": "gemini",
      "model_name": "gemini/gemini-3.1-flash-lite",
      "primary_failure_family": "none",
      "repair_suggestion": "No repair needed; this run demonstrates a robust successful policy for the task.",
      "run_uid": "gemini__gemini-3.1-flash-lite__browser_use__browser_use",
      "secondary_failure_tags": [],
      "success": true,
      "uncertainty_note": "The trace and contact sheet both show the success banner after save."
    }
  ],
  "schema_version": "layer2_task_v1",
  "task_id": "inline_editable_text-antd-T07",
  "task_level_judgment": {
    "benchmark_issue_note": "The success criterion appears consistent: the human and the two structured-access runs succeed only when the exact badge text, including punctuation, is entered. The failing runs appear to be agent-side near-misses rather than task flakiness.",
    "benchmark_issue_suspected": false,
    "confidence": "high",
    "cross_run_summary": "This task cleanly separates exact-text perception from UI manipulation. All agents discovered how to activate AntD inline editing and use the save control, but only the ax_tree and browser_use runs preserved the exact sample string with the em dash. The SOM and pixel runs visually approximated the label with a hyphen, reached a near-correct saved state, and then terminated anyway instead of checking for the promised automatic completion. That makes the task useful for measuring post-action verification under subtle visual-text mismatches.",
    "keep_revise_expand_retire_hint": "keep",
    "keep_revise_reason": "Keep this task because it is short, stable, and highly diagnostic of whether agents verify exact visual text matches rather than merely executing the editing workflow.",
    "overall_diagnostic_value": "high",
    "primary_task_story": "The central challenge in this task is not the AntD inline-edit interaction itself, which every run managed, but the requirement to match a visually shown badge label exactly, including a small punctuation distinction. Successful runs benefited from structured access to the exact string, while failing visual runs produced a semantically close but incorrect edit and then stopped without validating the task's automatic success condition.",
    "uncertainty_note": "The packet does not include raw screenshots for every intermediate state beyond the contact sheets, so the punctuation-reading error mechanism is inferred from the entered strings and final rendered states."
  }
}