| { |
| "canonical_type": "resizable_columns", |
| "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": "grounding", |
| "evidence_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" |
| ], |
| "evidence_step_refs": [ |
| "gemini__gemini-3.1-flash-lite__browsergym__ax_tree#0", |
| "gemini__gemini-3.1-flash-lite__browsergym__ax_tree#13", |
| "gemini__gemini-3.1-flash-lite__browsergym__som#3", |
| "gemini__gemini-3.1-flash-lite__browsergym__som#13", |
| "gemini__gemini-3.1-flash-lite__browsergym__pixel#0", |
| "gemini__gemini-3.1-flash-lite__browsergym__pixel#10", |
| "gemini__gemini-3.1-flash-lite__browser_use__browser_use#6", |
| "gemini__gemini-3.1-flash-lite__browser_use__browser_use#14" |
| ], |
| "likely_grounding_limited": true, |
| "likely_recovery_limited": true, |
| "likely_skill_fixable": true, |
| "secondary_failure_families": [ |
| "interaction_knowledge", |
| "recovery_state_tracking" |
| ], |
| "why": "The common bottleneck is acquiring and controlling the tiny MUI column separator at the right edge of the Status header. AX-tree never gets a clean resize-handle representation and drifts into menus; SOM and pixel can drag but hit neighboring affordances or wrong distances; browser_use knows the DOM structure but substitutes JS/event hacks for the required direct manipulation. After the first miss, runs rarely use the width monitor as tight feedback and instead loop through more drags, sorts, or menus." |
| }, |
| "family_id": "dragdrop_workspace", |
| "interesting_observation": "This task splits semantic understanding from motor control: ax_tree understands the column concept but cannot expose the separator, while pixel/SOM can move header boundaries yet still fail to stop at the requested 140px.", |
| "library": "mui", |
| "paper_useful_quote": "Although a human solved the task with one drag in 1.6 seconds, all four agent runs failed because the actionable target was a narrow MUI column separator adjacent to sort and menu affordances.", |
| "run_observations": [ |
| { |
| "confidence": "high", |
| "evidence_image_paths": [], |
| "evidence_step_ids": [ |
| 0, |
| 3, |
| 9, |
| 13, |
| 14 |
| ], |
| "failure_or_success_narrative": "AX-tree mode never acquires a usable resize handle. It begins with repeated `drag_and_drop('69','83')` attempts on the Status header region, then clicks header affordances, opens a column menu and later a column-visibility/search overlay, and finally reports infeasible because it cannot find a resize handle or width-setting option. The contact sheet shows sort/menu states rather than a controlled width change to 140px.", |
| "first_divergence_agent_step": 0, |
| "first_divergence_human_step": 0, |
| "framework": "browsergym", |
| "human_vs_agent_difference": "The human solved the task with one separator drag; this run treated the header and its menus as the actionable UI, producing sort/menu interactions instead of a resize.", |
| "is_benchmark_issue": false, |
| "is_grounding_limited": true, |
| "is_skill_fixable": true, |
| "mode": "ax_tree", |
| "model_family": "gemini", |
| "model_name": "gemini/gemini-3.1-flash-lite", |
| "primary_failure_family": "grounding", |
| "repair_suggestion": "Represent or heuristically infer the separator immediately to the right of the target header before exploring header menus, and use the width monitor as post-action feedback.", |
| "run_uid": "gemini__gemini-3.1-flash-lite__browsergym__ax_tree", |
| "secondary_failure_tags": [ |
| "small_target", |
| "did_not_know_ui_idiom", |
| "missed_overlay", |
| "cannot_recover_after_error" |
| ], |
| "success": false, |
| "uncertainty_note": "The exact semantics of AX nodes 69 and 83 are not decoded in the packet, but the screenshots clearly show menu/sort states instead of successful resizing." |
| }, |
| { |
| "confidence": "high", |
| "evidence_image_paths": [], |
| "evidence_step_ids": [ |
| 0, |
| 3, |
| 7, |
| 10, |
| 13, |
| 14 |
| ], |
| "failure_or_success_narrative": "SOM mode can physically drag something, but not precisely enough to control the Status separator. Early drags overshoot badly, with the contact sheet showing the status width reaching about 230px, after which the agent keeps dragging other header edges, clicks the header, opens the Status menu, and never converges back to 140px before max steps. The run demonstrates actuation without reliable target acquisition or calibration.", |
| "first_divergence_agent_step": 0, |
| "first_divergence_human_step": 0, |
| "framework": "browsergym", |
| "human_vs_agent_difference": "Instead of one calibrated drag to the target width, this run performed many exploratory drags and menu clicks, changing the column substantially but never stabilizing at 140px.", |
| "is_benchmark_issue": false, |
| "is_grounding_limited": true, |
| "is_skill_fixable": true, |
| "mode": "som", |
| "model_family": "gemini", |
| "model_name": "gemini/gemini-3.1-flash-lite", |
| "primary_failure_family": "grounding", |
| "repair_suggestion": "Once any width change is observed, switch to smaller corrective drags using the visible width monitor as closed-loop feedback rather than continuing large exploratory drags or opening menus.", |
| "run_uid": "gemini__gemini-3.1-flash-lite__browsergym__som", |
| "secondary_failure_tags": [ |
| "small_target", |
| "wrong_instance", |
| "cannot_recover_after_error", |
| "looping" |
| ], |
| "success": false, |
| "uncertainty_note": "The packet lists only the first 15 step summaries for a 20-step run, but the contact sheet already shows the critical overshoot and subsequent menu detour." |
| }, |
| { |
| "confidence": "high", |
| "evidence_image_paths": [], |
| "evidence_step_ids": [ |
| 0, |
| 5, |
| 10, |
| 12, |
| 14 |
| ], |
| "failure_or_success_narrative": "Pixel mode repeatedly executes horizontal drags near header boundaries, but it does not consistently anchor on the Status separator. The contact sheet shows attention shifting between the Status/Updated boundary and other nearby edges, with occasional sorting/highlighting side effects, yet no stable convergence to the required width. This is a classic tiny-target motor-control failure: raw pixels provide enough actuation to move things, but not enough reliable localization to hit the exact handle and stop at 140px.", |
| "first_divergence_agent_step": 1, |
| "first_divergence_human_step": 0, |
| "framework": "browsergym", |
| "human_vs_agent_difference": "The human completed a single precise drag; this run used many coordinate drags around adjacent header edges and never translated visual feedback into a correct final width.", |
| "is_benchmark_issue": false, |
| "is_grounding_limited": true, |
| "is_skill_fixable": true, |
| "mode": "pixel", |
| "model_family": "gemini", |
| "model_name": "gemini/gemini-3.1-flash-lite", |
| "primary_failure_family": "grounding", |
| "repair_suggestion": "Anchor the drag on the separator immediately to the right of the Status label and iteratively correct with smaller motions based on the width monitor after each drag.", |
| "run_uid": "gemini__gemini-3.1-flash-lite__browsergym__pixel", |
| "secondary_failure_tags": [ |
| "small_target", |
| "wrong_instance", |
| "looping", |
| "failed_to_confirm" |
| ], |
| "success": false, |
| "uncertainty_note": "Later steps are not fully enumerated in text, but the contact sheet and trace summary consistently show repeated drag attempts without convergence." |
| }, |
| { |
| "confidence": "medium", |
| "evidence_image_paths": [], |
| "evidence_step_ids": [ |
| 0, |
| 2, |
| 6, |
| 10, |
| 14 |
| ], |
| "failure_or_success_narrative": "Browser_use semantically understands the page and locates the Status header plus separator-like DOM nodes, but it never produces the one direct-manipulation drag that the component expects. Instead it loops through `evaluate(...)` calls to inspect the DOM, mutate styles, and synthesize drag events; those attempts do not reliably drive the visible widget to the requested 140px state, and the run times out. The failure is less about recognizing the target than about choosing an ineffective actuation strategy for this UI idiom.", |
| "first_divergence_agent_step": 0, |
| "first_divergence_human_step": 0, |
| "framework": "browser_use", |
| "human_vs_agent_difference": "The human performed one physical drag on the separator; this run stayed at the DOM/JS layer, attempting synthetic events and style hacks instead of a real pointer resize.", |
| "is_benchmark_issue": false, |
| "is_grounding_limited": false, |
| "is_skill_fixable": true, |
| "mode": "browser_use", |
| "model_family": "gemini", |
| "model_name": "gemini/gemini-3.1-flash-lite", |
| "primary_failure_family": "interaction_knowledge", |
| "repair_suggestion": "Prefer browser-level pointer interaction on the resize separator over DOM mutation or synthetic event injection, and verify success against the on-screen width monitor after each attempt.", |
| "run_uid": "gemini__gemini-3.1-flash-lite__browser_use__browser_use", |
| "secondary_failure_tags": [ |
| "did_not_know_ui_idiom", |
| "looping", |
| "failed_to_confirm", |
| "other" |
| ], |
| "success": false, |
| "uncertainty_note": "Only the first 15 of 21 steps are shown in the packet, so the exact final DOM state is partially truncated; the visible pattern is still a repeated evaluate loop rather than successful actuation." |
| } |
| ], |
| "schema_version": "layer2_task_v1", |
| "task_id": "resizable_columns-mui-T03", |
| "task_level_judgment": { |
| "benchmark_issue_note": "No benchmark issue is indicated by the packet. The human replay succeeded in one step, replay_success is true, and the failures cluster around separator acquisition and control rather than a broken success trigger or ambiguous instruction.", |
| "benchmark_issue_suspected": false, |
| "confidence": "high", |
| "cross_run_summary": "All four runs failed an instruction that the human completed with one drag because the actionable target is the narrow separator at the right edge of the Status column. AX-tree drifted into header menus because the resize handle was not cleanly exposed; SOM and pixel could drag but mis-targeted adjacent boundaries or overshot; browser_use identified the DOM structure but substituted JS/event hacks for a real drag. The visible width monitor existed, but agents did not use it as reliable closed-loop feedback.", |
| "keep_revise_expand_retire_hint": "revise", |
| "keep_revise_reason": "Keep the task because it isolates a real and common enterprise-grid failure mode, but revise its intended difficulty metadata upward: despite being labeled L0/easy, replay measurements and all agent runs show it behaves like a hard precision-control task.", |
| "overall_diagnostic_value": "high", |
| "primary_task_story": "This task is diagnostically strong because the instruction is simple, the scene is uncluttered, and the human solution is a single drag, yet agents still fail due to the combination of a tiny resize handle, nearby sort/menu affordances, and the need to stop at an exact width using live feedback.", |
| "uncertainty_note": "Human visual screenshots are not attached, so the analysis relies on the one-step human summary, alignment metadata, text traces, and the provided contact sheets for agent runs." |
| } |
| } |
