{
  "schema_version": "layer2_task_v1",
  "task_id": "tooltip-mantine-T08",
  "canonical_type": "tooltip",
  "family_id": "overlays_transient",
  "library": "mantine",
  "coverage": {
    "human_available": true,
    "human_visual_available": false,
    "realized_available": true,
    "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"
    ]
  },
  "task_level_judgment": {
    "overall_diagnostic_value": "high",
    "benchmark_issue_suspected": false,
    "benchmark_issue_note": "The human trace is unusually logged as a one-step scroll/no-op and the realized difficulty is low, but the supplied runs do not show a clearly broken checker or irreconcilable instruction mismatch.",
    "cross_run_summary": "This task cleanly separates observation modes. AX-tree solved it in one step by directly hovering an accessible node, browser_use solved it through DOM inspection and scripted targeting, pixel solved only after many exploratory scrolls plus a late cursor move into the panel, and SOM failed after looping on ineffective scrolls without ever exposing the Webhook URL row.",
    "primary_task_story": "Although framed as a tooltip task, the real challenge is first interacting with a nested scrollable settings card to reveal a lower row, then triggering a small info affordance. Modes with semantic access bypass that entirely, while purely visual control must discover the inner scroll context before the tooltip step even becomes possible.",
    "keep_revise_expand_retire_hint": "keep",
    "keep_revise_reason": "Despite low realized difficulty for semantic or DOM-aware agents, the task remains useful because it exposes whether agents understand nested scroll regions and can recover when repeated scrolls do not change visible state.",
    "confidence": "medium",
    "uncertainty_note": "Fine-grained human comparison is limited because the reference action is recorded as a scroll with zero delta, and success screenshots do not always visibly show the tooltip text at capture time."
  },
  "cross_run_analysis": {
    "dominant_failure_family": "interaction_knowledge",
    "secondary_failure_families": [
      "recovery_state_tracking"
    ],
    "likely_grounding_limited": false,
    "likely_skill_fixable": true,
    "likely_recovery_limited": true,
    "why": "The only outright failure is not a near-miss on the tooltip target; it is a failure to interact with the nested scrollable panel correctly. The SOM run kept issuing generic scrolls while the visible card stayed at the top rows, then gave up. In contrast, pixel eventually succeeded once it moved into the card before continuing, AX-tree bypassed the scroll problem by exposing the target semantically, and browser_use bypassed it through DOM queries. That pattern points to missing knowledge about sub-scrollable regions, with a secondary inability to stop and recover after unchanged state.",
    "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 steps 0-9",
      "gemini__gemini-3.1-flash-lite__browsergym__pixel steps 0-9",
      "gemini__gemini-3.1-flash-lite__browsergym__ax_tree step 0",
      "gemini__gemini-3.1-flash-lite__browser_use__browser_use steps 3-7"
    ]
  },
  "run_observations": [
    {
      "run_uid": "gemini__gemini-3.1-flash-lite__browsergym__ax_tree",
      "model_family": "gemini",
      "model_name": "gemini-3.1-flash-lite",
      "framework": "browsergym",
      "mode": "ax_tree",
      "success": true,
      "primary_failure_family": "none",
      "secondary_failure_tags": [],
      "human_vs_agent_difference": "Both trajectories finished immediately, but the agent used an explicit hover on AX node 65 rather than the human trace's recorded generic scroll action.",
      "failure_or_success_narrative": "The accessibility representation exposed the relevant control directly. At step 0 the agent hovered node `65`, and the task completed without any exploratory scrolling or visual search.",
      "repair_suggestion": "No repair needed in this mode; semantic exposure already makes the task trivial.",
      "is_skill_fixable": false,
      "is_grounding_limited": false,
      "is_benchmark_issue": false,
      "first_divergence_human_step": null,
      "first_divergence_agent_step": null,
      "evidence_step_ids": [
        0
      ],
      "evidence_image_paths": [],
      "confidence": "high",
      "uncertainty_note": "The reference trace uses a different recorded action type, so the exact equivalence to the human step is approximate."
    },
    {
      "run_uid": "gemini__gemini-3.1-flash-lite__browsergym__som",
      "model_family": "gemini",
      "model_name": "gemini-3.1-flash-lite",
      "framework": "browsergym",
      "mode": "som",
      "success": false,
      "primary_failure_family": "interaction_knowledge",
      "secondary_failure_tags": [
        "did_not_know_ui_idiom",
        "looping",
        "state_not_observed"
      ],
      "human_vs_agent_difference": "The human finished in one step, while this agent spent 10 steps issuing scrolls that never revealed the lower rows and never attempted a targeted hover inside the panel.",
      "failure_or_success_narrative": "Steps 0-8 are repeated `scroll(...)` actions, and the contact sheet keeps showing the same top-of-list rows such as Time zone, Locale, and Audit log. The agent never moved into the nested settings panel, never exposed `Webhook URL`, and then reported infeasible at step 9.",
      "repair_suggestion": "When repeated scrolls leave the panel unchanged, move the pointer into the scrollable card or target its scrollbar, then verify whether the inner content moved before continuing.",
      "is_skill_fixable": true,
      "is_grounding_limited": false,
      "is_benchmark_issue": false,
      "first_divergence_human_step": 0,
      "first_divergence_agent_step": 1,
      "evidence_step_ids": [
        0,
        1,
        2,
        3,
        4,
        5,
        6,
        7,
        8,
        9
      ],
      "evidence_image_paths": [],
      "confidence": "high",
      "uncertainty_note": "Not every intermediate cursor position is visible, but the panel contents shown in the contact sheet remain unchanged across the failed scroll attempts."
    },
    {
      "run_uid": "gemini__gemini-3.1-flash-lite__browsergym__pixel",
      "model_family": "gemini",
      "model_name": "gemini-3.1-flash-lite",
      "framework": "browsergym",
      "mode": "pixel",
      "success": true,
      "primary_failure_family": "none",
      "secondary_failure_tags": [
        "did_not_know_ui_idiom",
        "state_not_observed"
      ],
      "human_vs_agent_difference": "Instead of an immediate solution, the pixel agent needed several exploratory scrolls and a late cursor move into the card before it could expose and target the Webhook URL row.",
      "failure_or_success_narrative": "Steps 0-6 are mostly blind scroll attempts with little visible effect. After `mouse_move(640,360)` at step 7, the subsequent interaction finally scrolls the inner panel; by step 9 the cursor is over the `Webhook URL` area and the task completes, as shown in the final contact-sheet frame.",
      "repair_suggestion": "Detect the scrollable subpanel earlier and anchor scrolling to that region instead of repeating global scrolls.",
      "is_skill_fixable": true,
      "is_grounding_limited": true,
      "is_benchmark_issue": false,
      "first_divergence_human_step": 0,
      "first_divergence_agent_step": 1,
      "evidence_step_ids": [
        0,
        1,
        2,
        3,
        4,
        7,
        8,
        9
      ],
      "evidence_image_paths": [],
      "confidence": "high",
      "uncertainty_note": "The screenshots are sparse between actions, but the late appearance of the lower rows and scrollbar indicates the inner panel only started moving after the cursor entered it."
    },
    {
      "run_uid": "gemini__gemini-3.1-flash-lite__browser_use__browser_use",
      "model_family": "gemini",
      "model_name": "gemini-3.1-flash-lite",
      "framework": "browser_use",
      "mode": "browser_use",
      "success": true,
      "primary_failure_family": "none",
      "secondary_failure_tags": [],
      "human_vs_agent_difference": "Rather than directly performing the visible UI interaction, the agent navigated, inspected DOM structure, scrolled once, tried selector-based targeting, and then clicked the discovered trigger element.",
      "failure_or_success_narrative": "This run solved the task by bypassing perceptual ambiguity. After initial page inspection, it scrolled to the lower rows, used `evaluate(...)` probes to identify the Mantine tooltip trigger, and clicked element index 95 at step 7, which completed the task.",
      "repair_suggestion": "No repair is required; the main improvement would be reducing the extra selector-probing steps once the trigger is identified.",
      "is_skill_fixable": false,
      "is_grounding_limited": false,
      "is_benchmark_issue": false,
      "first_divergence_human_step": 0,
      "first_divergence_agent_step": 0,
      "evidence_step_ids": [
        0,
        1,
        3,
        4,
        5,
        6,
        7
      ],
      "evidence_image_paths": [],
      "confidence": "medium",
      "uncertainty_note": "The run finishes via click rather than an explicitly verified visible tooltip in the trace, though focus or click may still have opened the Mantine tooltip."
    }
  ],
  "interesting_observation": "The decisive move in the successful pixel run was not the final hover itself but the earlier cursor move into the settings card; before that, repeated scroll actions appeared to leave the nested list unchanged.",
  "paper_useful_quote": "This task separates semantic access from embodied control: the AX-tree agent finished in one hover, whereas the SOM agent looped on ineffective scrolls and never entered the nested scroll region."
}