{
  "imported_on": "2026-11-01",
  "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/",
  "task_count": 6,
  "trial_count": 14,
  "tasks": [
    {
      "id": "rcua_learn_blender_002",
      "kind": "blender",
      "name": {
        "en": "Blender · Exhibition stand",
        "zh": "Blender · 展台建模"
      },
      "version": "0.4.0-localvideo-en.1-saveguard.1",
      "language": "en",
      "application": "Blender",
      "instruction": "Prepare an editable exhibition stand for a fabrication review.\n\nBuild your own platform: dimensions 8 × 4 × 0.4 units. Place three round, closed-ended posts of diameter 1 along X at -2, 0 and 2. Their heights are 2, 3 and 4 respectively. The platform center can be anywhere; positions are relative to its center. All posts sit on its top face, at the same Y as the platform center. Keep the platform and posts independently editable (separate objects or disconnected mesh parts). Save before the first project pickup at 02:00.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
      "source_directory": "tasks/rcua_learn_blender_002",
      "download": "../downloads/rcua_learn_blender_002_en.zip",
      "tutorial": "evidence/blender/tutorial.mp4",
      "tutorial_sha256": "9c39daca9bcdad5ef3700e7664b13a151fb2a2eaa6afe3e7c50724355cbba0c8",
      "teacher_poster": "evidence/blender/teacher_poster.jpg",
      "runs": [
        {
          "id": "0",
          "label": "GPT‑6.1 Sol · local MP4 · 600s",
          "model": "GPT‑6.1 Sol",
          "score": 0.8,
          "version": "0.4.0-localvideo-en.1-saveguard.1",
          "model_run_version": "0.4.0-localvideo-en.1",
          "evaluation_version": "0.4.0-localvideo-en.1-saveguard.1",
          "budget_s": 600,
          "instruction": "Prepare an editable exhibition stand for a fabrication review.\n\nBuild your own platform: dimensions 8 × 4 × 0.4 units. Place three round, closed-ended posts of diameter 1 along X at -2, 0 and 2. Their heights are 2, 3 and 4 respectively. The platform center can be anywhere; positions are relative to its center. All posts sit on its top face, at the same Y as the platform center. Keep the platform and posts independently editable (separate objects or disconnected mesh parts). Save before the first project pickup at 02:00.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
          "video": "evidence/local/blender/gpt/model.mp4",
          "poster": "evidence/local/blender/gpt/poster.png",
          "result": "evidence/local/blender/gpt/result.json",
          "audit": "evidence/local/blender/gpt/publication_audit.json",
          "encoding": "evidence/local/blender/gpt/model.encoding.json",
          "sha256": "88371e5aa1ca02a6edbc2c624667de3b57fd6703a4c4815cb33777d8898856cf",
          "bytes": 847684,
          "partial_scores": {
            "workshop": 0,
            "final": 1.0
          },
          "regrade_note": "Saved native artifacts from the frozen model trial were regraded with the corrected evaluator. The public brief, starter, MP4, input protocol and clock are identical; this is not a new model execution.",
          "media_note": "Actual continuous native desktop recording, captured at 8 fps and encoded for the web at 3 fps. Playback preserves elapsed time. The agent switches between a normal local-MP4 player and the full-screen native application; instructions and client messages are plain text. No lesson dashboard or injected video frames.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/blender/gpt/model.mp4",
          "corrected_evaluation": "evidence/local/blender/gpt/corrected_evaluation.json",
          "original_runtime_source": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/blender_gpt_original_runtime_source.json",
          "download": "../downloads/rcua_learn_blender_002_en.zip",
          "source_directory": "tasks/rcua_learn_blender_002"
        },
        {
          "id": "1",
          "label": "Haiku 5.5 · local MP4 · 600s",
          "model": "Haiku 5.5",
          "score": 0.0,
          "version": "0.4.0-localvideo-en.1-saveguard.1",
          "model_run_version": "0.4.0-localvideo-en.1",
          "evaluation_version": "0.4.0-localvideo-en.1-saveguard.1",
          "budget_s": 600,
          "instruction": "Prepare an editable exhibition stand for a fabrication review.\n\nBuild your own platform: dimensions 8 × 4 × 0.4 units. Place three round, closed-ended posts of diameter 1 along X at -2, 0 and 2. Their heights are 2, 3 and 4 respectively. The platform center can be anywhere; positions are relative to its center. All posts sit on its top face, at the same Y as the platform center. Keep the platform and posts independently editable (separate objects or disconnected mesh parts). Save before the first project pickup at 02:00.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
          "video": "evidence/local/blender/haiku/model.mp4",
          "poster": "evidence/local/blender/haiku/poster.png",
          "result": "evidence/local/blender/haiku/result.json",
          "audit": "evidence/local/blender/haiku/publication_audit.json",
          "encoding": "evidence/local/blender/haiku/model.encoding.json",
          "sha256": "9fa40c68afc2670fa108743c50d3705428421b493c3cd2a39ea0abbd8bb1d3d7",
          "bytes": 411027,
          "partial_scores": {
            "workshop": 0,
            "final": 0
          },
          "regrade_note": "Saved native artifacts from the frozen model trial were regraded with the corrected evaluator. The public brief, starter, MP4, input protocol and clock are identical; this is not a new model execution.",
          "media_note": "Actual continuous native desktop recording, captured at 8 fps and encoded for the web at 3 fps. Playback preserves elapsed time. The agent switches between a normal local-MP4 player and the full-screen native application; instructions and client messages are plain text. No lesson dashboard or injected video frames.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/blender/haiku/model.mp4",
          "corrected_evaluation": "evidence/local/blender/haiku/corrected_evaluation.json",
          "original_runtime_source": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/blender_haiku_original_runtime_source.json",
          "download": "../downloads/rcua_learn_blender_002_en.zip",
          "source_directory": "tasks/rcua_learn_blender_002"
        }
      ]
    },
    {
      "id": "rcua_learn_calc_001",
      "kind": "calc",
      "name": {
        "en": "Calc · Video-only settlement",
        "zh": "Calc · 纯视频结算规则"
      },
      "version": "0.7.0-video-only-self-notes-en.1",
      "language": "en",
      "application": "LibreOffice Calc",
      "instruction": "Prepare Orders.ods for the operations handoff demonstrated in the live walkthrough. Apply the demonstrated settlement policy to your supplied orders, incorporate any correction announced during the recording, and deliver an editable workbook. The demonstration uses different sample orders.\n\nThe supplied local Tutorial.mp4 is already open in an ordinary full-screen player. Business rules, required workbook changes and their revisions are communicated only inside that video; no text transcript or chat correction will be supplied. You may start working immediately and switch between the player and Calc with Alt+Tab. The video plays once at normal speed and continues while Calc is foreground. Pause, seek, replay and speed changes are disabled. Only pointer movement and Alt+Tab are accepted while the player is active; rejected inputs consume a step.\n\nA copy of your latest saved workbook will be collected at 02:00. The final handoff is due at 10:00 from task start, regardless of which window is active. Preserve supplied input cells and use the normal Save command for Orders.ods. A screenshot or completion message is not a saved workbook.\n\nUse screenshots and ordinary mouse/keyboard input only. Native spreadsheet formulas and application commands are allowed. Do not use terminals, scripts, macros, developer tools, APIs, project-source editing or private reference files.\n\nYou may retain brief factual notes from your own observations in the tool memory field. This field is not GUI input. No notes, transcript or business parameters will be supplied by the harness.",
      "source_directory": "tasks/rcua_learn_calc_001",
      "download": "../downloads/rcua_learn_calc_001_en.zip",
      "tutorial": "evidence/video-only/tutorial.mp4",
      "tutorial_sha256": "0ea4b78d32be191cdfa9d3addcc9bf0a7f28e3cf0fcc192738f00029394d6a79",
      "teacher_poster": "evidence/video-self-notes/teacher_poster.webp",
      "runs": [
        {
          "id": "0",
          "label": "GPT‑6.1 Sol · video-only + self-notes ·600s",
          "model": "GPT‑6.1 Sol",
          "score": 1.0,
          "version": "0.7.0-video-only-self-notes-en.1",
          "model_run_version": "0.7.0-video-only-self-notes-en.1",
          "evaluation_version": "0.7.0-video-only-self-notes-en.1",
          "budget_s": 600,
          "instruction": "Prepare Orders.ods for the operations handoff demonstrated in the live walkthrough. Apply the demonstrated settlement policy to your supplied orders, incorporate any correction announced during the recording, and deliver an editable workbook. The demonstration uses different sample orders.\n\nThe supplied local Tutorial.mp4 is already open in an ordinary full-screen player. Business rules, required workbook changes and their revisions are communicated only inside that video; no text transcript or chat correction will be supplied. You may start working immediately and switch between the player and Calc with Alt+Tab. The video plays once at normal speed and continues while Calc is foreground. Pause, seek, replay and speed changes are disabled. Only pointer movement and Alt+Tab are accepted while the player is active; rejected inputs consume a step.\n\nA copy of your latest saved workbook will be collected at 02:00. The final handoff is due at 10:00 from task start, regardless of which window is active. Preserve supplied input cells and use the normal Save command for Orders.ods. A screenshot or completion message is not a saved workbook.\n\nUse screenshots and ordinary mouse/keyboard input only. Native spreadsheet formulas and application commands are allowed. Do not use terminals, scripts, macros, developer tools, APIs, project-source editing or private reference files.\n\nYou may retain brief factual notes from your own observations in the tool memory field. This field is not GUI input. No notes, transcript or business parameters will be supplied by the harness.",
          "video": "evidence/video-self-notes/gpt/model.mp4",
          "poster": "evidence/video-self-notes/gpt/poster.webp",
          "result": "evidence/video-self-notes/gpt/result.json",
          "audit": "evidence/video-self-notes/gpt/publication_audit.json",
          "encoding": "evidence/video-self-notes/gpt/encoding.json",
          "sha256": "8009661edf2d90464eb3a0349f5c695229d63c407b6ee0be898ea1105494595c",
          "bytes": 926326,
          "partial_scores": {
            "workshop": 1.0,
            "final": 1.0
          },
          "regrade_note": null,
          "media_note": "Actual continuous8fps native desktop capture, web-encoded at3fps with the full elapsed timeline. Task playback was one-pass1x with no pause/seek; the web replay is freely seekable. Business rules and changes appeared only inside the video; no text policy messages or teacher-frame injections.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/video-self-notes/gpt/model.mp4",
          "original_runtime_source": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/video-self-notes/original_runtime_source.json",
          "download": "../downloads/rcua_learn_calc_001_en.zip",
          "source_directory": "tasks/rcua_learn_calc_001"
        },
        {
          "id": "1",
          "label": "Haiku 5.5 · video-only + self-notes ·600s",
          "model": "Haiku 5.5",
          "score": 0.0,
          "version": "0.7.0-video-only-self-notes-en.1",
          "model_run_version": "0.7.0-video-only-self-notes-en.1",
          "evaluation_version": "0.7.0-video-only-self-notes-en.1",
          "budget_s": 600,
          "instruction": "Prepare Orders.ods for the operations handoff demonstrated in the live walkthrough. Apply the demonstrated settlement policy to your supplied orders, incorporate any correction announced during the recording, and deliver an editable workbook. The demonstration uses different sample orders.\n\nThe supplied local Tutorial.mp4 is already open in an ordinary full-screen player. Business rules, required workbook changes and their revisions are communicated only inside that video; no text transcript or chat correction will be supplied. You may start working immediately and switch between the player and Calc with Alt+Tab. The video plays once at normal speed and continues while Calc is foreground. Pause, seek, replay and speed changes are disabled. Only pointer movement and Alt+Tab are accepted while the player is active; rejected inputs consume a step.\n\nA copy of your latest saved workbook will be collected at 02:00. The final handoff is due at 10:00 from task start, regardless of which window is active. Preserve supplied input cells and use the normal Save command for Orders.ods. A screenshot or completion message is not a saved workbook.\n\nUse screenshots and ordinary mouse/keyboard input only. Native spreadsheet formulas and application commands are allowed. Do not use terminals, scripts, macros, developer tools, APIs, project-source editing or private reference files.\n\nYou may retain brief factual notes from your own observations in the tool memory field. This field is not GUI input. No notes, transcript or business parameters will be supplied by the harness.",
          "video": "evidence/video-self-notes/haiku/model.mp4",
          "poster": "evidence/video-self-notes/haiku/poster.webp",
          "result": "evidence/video-self-notes/haiku/result.json",
          "audit": "evidence/video-self-notes/haiku/publication_audit.json",
          "encoding": "evidence/video-self-notes/haiku/encoding.json",
          "sha256": "38c91841f3a6e8da5c3acc01c4403ae8304f26dfb7e7b40568efee0131c9bca9",
          "bytes": 733447,
          "partial_scores": {
            "workshop": 0,
            "final": 0
          },
          "regrade_note": null,
          "media_note": "Actual continuous8fps native desktop capture, web-encoded at3fps with the full elapsed timeline. Task playback was one-pass1x with no pause/seek; the web replay is freely seekable. Business rules and changes appeared only inside the video; no text policy messages or teacher-frame injections.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/video-self-notes/haiku/model.mp4",
          "original_runtime_source": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/video-self-notes/original_runtime_source.json",
          "download": "../downloads/rcua_learn_calc_001_en.zip",
          "source_directory": "tasks/rcua_learn_calc_001"
        }
      ]
    },
    {
      "id": "rcua_learn_impress_001",
      "kind": "impress",
      "name": {
        "en": "Impress · Brand one-pager",
        "zh": "Impress · 品牌单页"
      },
      "version": "0.4.0-localvideo-en.15",
      "language": "en",
      "application": "LibreOffice Impress",
      "instruction": "Prepare an editable brand one-pager for the client.\n\nUpdate your one-pager: Harbor Metrics → Northline Labs; Operations snapshot → Field report; Quarterly review → Service operations. Set CONVERSION 24%, RESPONSE 28m, RETENTION 96%. Change Draft - internal to Working draft. Keep every element editable. Save before the first project pickup at 02:00.\n\nKeep one visible card per metric: show each metric label once, paired with one value. Replace old brand and approval text completely, and keep all required text legible at normal slide size.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
      "source_directory": "tasks/rcua_learn_impress_001",
      "download": "../downloads/rcua_learn_impress_001_en.zip",
      "tutorial": "evidence/impress/tutorial.mp4",
      "tutorial_sha256": "ba3ef5ab49941cf9f538dfbd1a4e063b61f7cb4005a3d430c28d1f527e969c7a",
      "teacher_poster": "evidence/impress/teacher_poster.jpg",
      "runs": [
        {
          "id": "0",
          "label": "GPT‑6.1 Sol · local MP4 · 600s",
          "model": "GPT‑6.1 Sol",
          "score": 0.8,
          "version": "0.4.0-localvideo-en.15",
          "model_run_version": "0.4.0-localvideo-en.12",
          "evaluation_version": "0.4.0-localvideo-en.15",
          "budget_s": 600,
          "instruction": "Prepare an editable brand one-pager for the client.\n\nUpdate your one-pager: Harbor Metrics → Northline Labs; Operations snapshot → Field report; Quarterly review → Service operations. Set CONVERSION 24%, RESPONSE 28m, RETENTION 96%. Change Draft - internal to Working draft. Keep every element editable. Save before the first project pickup at 02:00.\n\nKeep one visible card per metric: show each metric label once, paired with one value. Replace old brand and approval text completely, and keep all required text legible at normal slide size.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
          "video": "evidence/local/impress/gpt/model.mp4",
          "poster": "evidence/local/impress/gpt/poster.png",
          "result": "evidence/local/impress/gpt/result.json",
          "audit": "evidence/local/impress/gpt/publication_audit.json",
          "encoding": "evidence/local/impress/gpt/model.encoding.json",
          "sha256": "9515835e0c74edbcdfe367a55e56677e8d0b7da62a43a0a0662a14b29abf60c1",
          "bytes": 554191,
          "partial_scores": {
            "workshop": 0,
            "final": 1.0
          },
          "regrade_note": "Saved native artifacts from the frozen model trial were regraded with the corrected evaluator. The public brief, starter, MP4, input protocol and clock are identical; this is not a new model execution.",
          "media_note": "Actual continuous native desktop recording, captured at 8 fps and encoded for the web at 3 fps. Playback preserves elapsed time. The agent switches between a normal local-MP4 player and the full-screen native application; instructions and client messages are plain text. No lesson dashboard or injected video frames.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/impress/gpt/model.mp4",
          "corrected_evaluation": "evidence/local/impress/gpt/corrected_evaluation.json",
          "original_runtime_source": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/impress_gpt_original_runtime_source.json",
          "download": "../downloads/rcua_learn_impress_001_en.zip",
          "source_directory": "tasks/rcua_learn_impress_001"
        },
        {
          "id": "1",
          "label": "Haiku 5.5 · local MP4 · 600s",
          "model": "Haiku 5.5",
          "score": 0.64,
          "version": "0.4.0-localvideo-en.15",
          "model_run_version": "0.4.0-localvideo-en.12",
          "evaluation_version": "0.4.0-localvideo-en.15",
          "budget_s": 600,
          "instruction": "Prepare an editable brand one-pager for the client.\n\nUpdate your one-pager: Harbor Metrics → Northline Labs; Operations snapshot → Field report; Quarterly review → Service operations. Set CONVERSION 24%, RESPONSE 28m, RETENTION 96%. Change Draft - internal to Working draft. Keep every element editable. Save before the first project pickup at 02:00.\n\nKeep one visible card per metric: show each metric label once, paired with one value. Replace old brand and approval text completely, and keep all required text legible at normal slide size.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
          "video": "evidence/local/impress/haiku/model.mp4",
          "poster": "evidence/local/impress/haiku/poster.png",
          "result": "evidence/local/impress/haiku/result.json",
          "audit": "evidence/local/impress/haiku/publication_audit.json",
          "encoding": "evidence/local/impress/haiku/model.encoding.json",
          "sha256": "370e56254597b8fa304a78395aa5233caddcdd400a5c61d88ee5026361fd9c5a",
          "bytes": 568082,
          "partial_scores": {
            "workshop": 0,
            "final": 0.8
          },
          "regrade_note": "Saved native artifacts from the frozen model trial were regraded with the corrected evaluator. The public brief, starter, MP4, input protocol and clock are identical; this is not a new model execution.",
          "media_note": "Actual continuous native desktop recording, captured at 8 fps and encoded for the web at 3 fps. Playback preserves elapsed time. The agent switches between a normal local-MP4 player and the full-screen native application; instructions and client messages are plain text. No lesson dashboard or injected video frames.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/impress/haiku/model.mp4",
          "corrected_evaluation": "evidence/local/impress/haiku/corrected_evaluation.json",
          "original_runtime_source": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/impress_haiku_original_runtime_source.json",
          "download": "../downloads/rcua_learn_impress_001_en.zip",
          "source_directory": "tasks/rcua_learn_impress_001"
        }
      ]
    },
    {
      "id": "rcua_learn_inkscape_001",
      "kind": "inkscape",
      "name": {
        "en": "Inkscape · Wayfinding sign",
        "zh": "Inkscape · 导视标识"
      },
      "version": "0.4.0-localvideo-en.2-saveguard.1",
      "language": "en",
      "application": "Inkscape",
      "instruction": "Prepare an editable wayfinding sign for the venue team.\n\nCreate one orange editable path from the two supplied circles and arrowhead. The ring center is (150,100) px, outer diameter 120 px and transparent inner diameter 72 px. The arrowhead points right, from (194,70) through (230,100) to (194,130). Use orange #E67E45. Keep the hole transparent. Set the editable caption to OPEN STUDIO. Save Waypoint.svg before the first project pickup at 02:00.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
      "source_directory": "tasks/rcua_learn_inkscape_001",
      "download": "../downloads/rcua_learn_inkscape_001_en.zip",
      "tutorial": "evidence/inkscape/tutorial.mp4",
      "tutorial_sha256": "d6fa8b93662a5bd655946f00f903bf8620492290a042ee618f9d206e61200351",
      "teacher_poster": "evidence/inkscape/teacher_poster.jpg",
      "runs": [
        {
          "id": "0",
          "label": "GPT‑6.1 Sol · local MP4 · 600s",
          "model": "GPT‑6.1 Sol",
          "score": 0.8,
          "version": "0.4.0-localvideo-en.2-saveguard.1",
          "model_run_version": "0.4.0-localvideo-en.2",
          "evaluation_version": "0.4.0-localvideo-en.2-saveguard.1",
          "budget_s": 600,
          "instruction": "Prepare an editable wayfinding sign for the venue team.\n\nCreate one orange editable path from the two supplied circles and arrowhead. The ring center is (150,100) px, outer diameter 120 px and transparent inner diameter 72 px. The arrowhead points right, from (194,70) through (230,100) to (194,130). Use orange #E67E45. Keep the hole transparent. Set the editable caption to OPEN STUDIO. Save Waypoint.svg before the first project pickup at 02:00.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
          "video": "evidence/local/inkscape/gpt/model.mp4",
          "poster": "evidence/local/inkscape/gpt/poster.png",
          "result": "evidence/local/inkscape/gpt/result.json",
          "audit": "evidence/local/inkscape/gpt/publication_audit.json",
          "encoding": "evidence/local/inkscape/gpt/model.encoding.json",
          "sha256": "e1760996e0a6bb6f5cbbd4374645ba9ed6f5968bfc3c043decd17d4f626e2ece",
          "bytes": 619590,
          "partial_scores": {
            "workshop": 0,
            "final": 1.0
          },
          "regrade_note": "Saved native artifacts from the frozen model trial were regraded with the corrected evaluator. The public brief, starter, MP4, input protocol and clock are identical; this is not a new model execution.",
          "media_note": "Actual continuous native desktop recording, captured at 8 fps and encoded for the web at 3 fps. Playback preserves elapsed time. The agent switches between a normal local-MP4 player and the full-screen native application; instructions and client messages are plain text. No lesson dashboard or injected video frames.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/inkscape/gpt/model.mp4",
          "corrected_evaluation": "evidence/local/inkscape/gpt/corrected_evaluation.json",
          "original_runtime_source": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/inkscape_gpt_original_runtime_source.json",
          "download": "../downloads/rcua_learn_inkscape_001_en.zip",
          "source_directory": "tasks/rcua_learn_inkscape_001"
        },
        {
          "id": "1",
          "label": "Haiku 5.5 · local MP4 · 600s",
          "model": "Haiku 5.5",
          "score": 0.0,
          "version": "0.4.0-localvideo-en.2-saveguard.1",
          "model_run_version": "0.4.0-localvideo-en.2",
          "evaluation_version": "0.4.0-localvideo-en.2-saveguard.1",
          "budget_s": 600,
          "instruction": "Prepare an editable wayfinding sign for the venue team.\n\nCreate one orange editable path from the two supplied circles and arrowhead. The ring center is (150,100) px, outer diameter 120 px and transparent inner diameter 72 px. The arrowhead points right, from (194,70) through (230,100) to (194,130). Use orange #E67E45. Keep the hole transparent. Set the editable caption to OPEN STUDIO. Save Waypoint.svg before the first project pickup at 02:00.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
          "video": "evidence/local/inkscape/haiku/model.mp4",
          "poster": "evidence/local/inkscape/haiku/poster.png",
          "result": "evidence/local/inkscape/haiku/result.json",
          "audit": "evidence/local/inkscape/haiku/publication_audit.json",
          "encoding": "evidence/local/inkscape/haiku/model.encoding.json",
          "sha256": "49701f7bca8318415afee07adeae32c0223633bad0f069109e91bda084453b8c",
          "bytes": 366789,
          "partial_scores": {
            "workshop": 0,
            "final": 0
          },
          "regrade_note": "Saved native artifacts from the frozen model trial were regraded with the corrected evaluator. The public brief, starter, MP4, input protocol and clock are identical; this is not a new model execution.",
          "media_note": "Actual continuous native desktop recording, captured at 8 fps and encoded for the web at 3 fps. Playback preserves elapsed time. The agent switches between a normal local-MP4 player and the full-screen native application; instructions and client messages are plain text. No lesson dashboard or injected video frames.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/inkscape/haiku/model.mp4",
          "corrected_evaluation": "evidence/local/inkscape/haiku/corrected_evaluation.json",
          "original_runtime_source": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/inkscape_haiku_original_runtime_source.json",
          "download": "../downloads/rcua_learn_inkscape_001_en.zip",
          "source_directory": "tasks/rcua_learn_inkscape_001"
        }
      ]
    },
    {
      "id": "rcua_learn_kicad_001",
      "kind": "kicad",
      "name": {
        "en": "KiCad · Signal routing",
        "zh": "KiCad · 电路板走线"
      },
      "version": "0.4.0-localvideo-en.2-saveguard.1",
      "language": "en",
      "application": "KiCad PCB Editor",
      "instruction": "Prepare the signal board for an engineering review.\n\nPlace J1 at (10,15), R1 at (25,15) and D1 at (40,15) mm, with zero rotation. Complete all three nets: PWR from J1 pad 1 to R1 pad 1; LED from R1 pad 2 to D1 pad 1; GND from D1 pad 2 to J1 pad 2. Keep the supplied 50×30 mm outline and four mounting holes. Use copper tracks at least 0.2 mm wide. Save the first routed board before the first project pickup at 02:00.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
      "source_directory": "tasks/rcua_learn_kicad_001",
      "download": "../downloads/rcua_learn_kicad_001_en.zip",
      "tutorial": "evidence/kicad/tutorial.mp4",
      "tutorial_sha256": "031e18c22619a6ef6026ffd39f3a0810052d08688c9c249db5326c3bd4e3f7c7",
      "teacher_poster": "evidence/kicad/teacher_poster.jpg",
      "runs": [
        {
          "id": "0",
          "label": "GPT‑6.1 Sol · local MP4 · 600s",
          "model": "GPT‑6.1 Sol",
          "score": 0.8,
          "version": "0.4.0-localvideo-en.2-saveguard.1",
          "model_run_version": "0.4.0-localvideo-en.2",
          "evaluation_version": "0.4.0-localvideo-en.2-saveguard.1",
          "budget_s": 600,
          "instruction": "Prepare the signal board for an engineering review.\n\nPlace J1 at (10,15), R1 at (25,15) and D1 at (40,15) mm, with zero rotation. Complete all three nets: PWR from J1 pad 1 to R1 pad 1; LED from R1 pad 2 to D1 pad 1; GND from D1 pad 2 to J1 pad 2. Keep the supplied 50×30 mm outline and four mounting holes. Use copper tracks at least 0.2 mm wide. Save the first routed board before the first project pickup at 02:00.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
          "video": "evidence/local/kicad/gpt/model.mp4",
          "poster": "evidence/local/kicad/gpt/poster.png",
          "result": "evidence/local/kicad/gpt/result.json",
          "audit": "evidence/local/kicad/gpt/publication_audit.json",
          "encoding": "evidence/local/kicad/gpt/model.encoding.json",
          "sha256": "4ed54a24d7691ac990f6c6613e41aebff6dc519006f20b51ba263291ab3598f3",
          "bytes": 1474494,
          "partial_scores": {
            "workshop": 0,
            "final": 1.0
          },
          "regrade_note": "Saved native artifacts from the frozen model trial were regraded with the corrected evaluator. The public brief, starter, MP4, input protocol and clock are identical; this is not a new model execution.",
          "media_note": "Actual continuous native desktop recording, captured at 8 fps and encoded for the web at 3 fps. Playback preserves elapsed time. The agent switches between a normal local-MP4 player and the full-screen native application; instructions and client messages are plain text. No lesson dashboard or injected video frames.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/kicad/gpt/model.mp4",
          "corrected_evaluation": "evidence/local/kicad/gpt/corrected_evaluation.json",
          "original_runtime_source": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/kicad_gpt_original_runtime_source.json",
          "download": "../downloads/rcua_learn_kicad_001_en.zip",
          "source_directory": "tasks/rcua_learn_kicad_001"
        },
        {
          "id": "1",
          "label": "Haiku 5.5 · local MP4 · 600s",
          "model": "Haiku 5.5",
          "score": 0.0,
          "version": "0.4.0-localvideo-en.2-saveguard.1",
          "model_run_version": "0.4.0-localvideo-en.2",
          "evaluation_version": "0.4.0-localvideo-en.2-saveguard.1",
          "budget_s": 600,
          "instruction": "Prepare the signal board for an engineering review.\n\nPlace J1 at (10,15), R1 at (25,15) and D1 at (40,15) mm, with zero rotation. Complete all three nets: PWR from J1 pad 1 to R1 pad 1; LED from R1 pad 2 to D1 pad 1; GND from D1 pad 2 to J1 pad 2. Keep the supplied 50×30 mm outline and four mounting holes. Use copper tracks at least 0.2 mm wide. Save the first routed board before the first project pickup at 02:00.\n\nUse the open native application and the supplied Tutorial.mp4, which is already open in the ordinary video player. The tutorial shows a different example: learn the operations and apply them to this project. You may start working immediately, switch applications with Alt+Tab, and pause or seek the video with its normal controls. There is no lesson website or review button.\n\nThe design lead collects a copy of your latest saved working file after 2 minutes. Later edits cannot change that copy. Finish the final handoff within 10 minutes of starting; pausing the video does not extend either deadline. Save with the application's normal Save command. Client corrections may arrive as text messages while you work; follow the latest request.\n\nOperate with screenshots and ordinary mouse/keyboard input only. Numeric fields, spreadsheet formulas and native GUI commands are allowed. Do not use terminals, application scripting consoles, macros, developer tools, APIs, project-source editing or private reference files. Keep the deliverable editable and preserve supplied input data. A screenshot or a completion message does not save the native project.",
          "video": "evidence/local/kicad/haiku/model.mp4",
          "poster": "evidence/local/kicad/haiku/poster.png",
          "result": "evidence/local/kicad/haiku/result.json",
          "audit": "evidence/local/kicad/haiku/publication_audit.json",
          "encoding": "evidence/local/kicad/haiku/model.encoding.json",
          "sha256": "54e5e4441fb2689e005ecd087365d825d7c011d23fcd03fcfda2d403867b2573",
          "bytes": 731241,
          "partial_scores": {
            "workshop": 0,
            "final": 0
          },
          "regrade_note": "Saved native artifacts from the frozen model trial were regraded with the corrected evaluator. The public brief, starter, MP4, input protocol and clock are identical; this is not a new model execution.",
          "media_note": "Actual continuous native desktop recording, captured at 8 fps and encoded for the web at 3 fps. Playback preserves elapsed time. The agent switches between a normal local-MP4 player and the full-screen native application; instructions and client messages are plain text. No lesson dashboard or injected video frames.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/kicad/haiku/model.mp4",
          "corrected_evaluation": "evidence/local/kicad/haiku/corrected_evaluation.json",
          "original_runtime_source": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/local/kicad_haiku_original_runtime_source.json",
          "download": "../downloads/rcua_learn_kicad_001_en.zip",
          "source_directory": "tasks/rcua_learn_kicad_001"
        }
      ]
    },
    {
      "id": "rcua_learn_godot_002",
      "kind": "godot",
      "name": {
        "en": "Godot · Bridge route",
        "zh": "Godot · 桥梁路线"
      },
      "version": "0.2.1-fullscreen-budget600-en.1",
      "language": "en",
      "application": "Godot",
      "instruction": "Work alongside the continuous native-software demonstration. The instructor works on a different example; apply the method to your own open project. You may edit while watching. Playback runs for 180 seconds without pausing or seeking. The final deadline is 420 seconds after playback ends, at 10:00. Save a first draft before the 02:00 workshop collection. That immutable draft is worth 20%; final saved content is worth 80%. Later work cannot recover a missed workshop draft. Partial credit reflects the published components of the actual native project.\n\nUse screenshots and ordinary mouse/keyboard only. No terminals, application scripting consoles, scripts, macros, developer tools, DOM, HTTP calls or editing project source as text. Normal numeric fields, object properties, text labels and native application commands are allowed. The task page has optional technique help and saved-work reviews from a scripted instructor. Help and inference consume wall time. The instructor is simulated, not a human. Instructions and feedback refer to released requirements; a later revision overrides earlier advice. Saving or claiming completion on the task page does not edit or save the native application.\n\nSave the editable project in its existing workspace file. Captures use trusted observation time, not file timestamps. The last captured native file before each collection is used; missing or corrupt final delivery earns zero overall. Runtime or reference failures are reported separately. No symbolic links; maximum file size 64 MiB. Equivalent editable constructions and object renaming are accepted unless a reference label identifies a circuit component. Do not replace the project with a screenshot.\n\nDeliver Harbor.tscn in the supplied project, preserving the supplied read-only project.godot. Each collection scores: two correctly positioned visible bridges 15% each; two matching real box collisions 15% each; continuous collision-supported connection between the fixed platforms 20%; unchanged fixed platforms 10%; current lane/clearance compliance with exactly two bridge pieces 10%. Collision credit is conditional on the corresponding bridge reaching its requested position. Position and dimension tolerance 0.05 m; collision/visual mismatch tolerance 0.02 m. Keep exactly the four supplied floor bodies with one BoxMesh surface and one direct BoxShape3D collision each; do not add floor geometry. Preserve their collision and render layers. Keep the supplied BoxMesh and BoxShape3D primitives; node names may change. No custom scripts, external scene assets, executable resources or replacement flattened images; keep a native editable Node3D scene with ordinary MeshInstance3D and CollisionShape3D nodes.\n\nFULL-SCREEN WINDOW-SWITCHING PILOT: The Lesson Player and Godot editor are two independent FULL-SCREEN applications. Only the front application is visible. When Godot is in front, the complete player, its lesson notes and its video are hidden. There is NO persistent sidebar or overlay. Use Alt+Tab to switch between the real editor and the player. Clicking controls pictured inside the recorded video never edits your project. Keep both applications full screen.\nWatch the operation video, switch to the real editor, apply the technique to your supplied different project and save with Ctrl+S. Return to the player after editing. Revisit near02:05 to check the revised brief and the instructor's revised operation, then return to your project. Playback, requirement changes and the final deadline continue while the player is hidden. No background tutorial contact sheets, hidden frames, pause/seek controls or automatic window switching are supplied. Native content scores and observed switching/full-screen compliance are reported separately.",
      "source_directory": "tasks/rcua_learn_godot_002",
      "download": "../downloads/rcua_learn_godot_002_en.zip",
      "tutorial": "evidence/godot/tutorial.mp4",
      "tutorial_sha256": "4a402e2196ae3350be8ade976e9678dbc90a9d337f717fa77081a05387ef04b4",
      "teacher_poster": "evidence/godot/teacher_poster.jpg",
      "runs": [
        {
          "id": "0",
          "label": "GPT‑6.1 Sol · 600s",
          "model": "GPT‑6.1 Sol",
          "score": 0.82,
          "version": "0.2.1-fullscreen-budget600-en.1",
          "model_run_version": "0.2.1-fullscreen-budget600-en.1",
          "evaluation_version": "0.2.1-fullscreen-budget600-en.1",
          "budget_s": 600,
          "instruction": "Work alongside the continuous native-software demonstration. The instructor works on a different example; apply the method to your own open project. You may edit while watching. Playback runs for 180 seconds without pausing or seeking. The final deadline is 420 seconds after playback ends, at 10:00. Save a first draft before the 02:00 workshop collection. That immutable draft is worth 20%; final saved content is worth 80%. Later work cannot recover a missed workshop draft. Partial credit reflects the published components of the actual native project.\n\nUse screenshots and ordinary mouse/keyboard only. No terminals, application scripting consoles, scripts, macros, developer tools, DOM, HTTP calls or editing project source as text. Normal numeric fields, object properties, text labels and native application commands are allowed. The task page has optional technique help and saved-work reviews from a scripted instructor. Help and inference consume wall time. The instructor is simulated, not a human. Instructions and feedback refer to released requirements; a later revision overrides earlier advice. Saving or claiming completion on the task page does not edit or save the native application.\n\nSave the editable project in its existing workspace file. Captures use trusted observation time, not file timestamps. The last captured native file before each collection is used; missing or corrupt final delivery earns zero overall. Runtime or reference failures are reported separately. No symbolic links; maximum file size 64 MiB. Equivalent editable constructions and object renaming are accepted unless a reference label identifies a circuit component. Do not replace the project with a screenshot.\n\nDeliver Harbor.tscn in the supplied project, preserving the supplied read-only project.godot. Each collection scores: two correctly positioned visible bridges 15% each; two matching real box collisions 15% each; continuous collision-supported connection between the fixed platforms 20%; unchanged fixed platforms 10%; current lane/clearance compliance with exactly two bridge pieces 10%. Collision credit is conditional on the corresponding bridge reaching its requested position. Position and dimension tolerance 0.05 m; collision/visual mismatch tolerance 0.02 m. Keep exactly the four supplied floor bodies with one BoxMesh surface and one direct BoxShape3D collision each; do not add floor geometry. Preserve their collision and render layers. Keep the supplied BoxMesh and BoxShape3D primitives; node names may change. No custom scripts, external scene assets, executable resources or replacement flattened images; keep a native editable Node3D scene with ordinary MeshInstance3D and CollisionShape3D nodes.\n\nFULL-SCREEN WINDOW-SWITCHING PILOT: The Lesson Player and Godot editor are two independent FULL-SCREEN applications. Only the front application is visible. When Godot is in front, the complete player, its lesson notes and its video are hidden. There is NO persistent sidebar or overlay. Use Alt+Tab to switch between the real editor and the player. Clicking controls pictured inside the recorded video never edits your project. Keep both applications full screen.\nWatch the operation video, switch to the real editor, apply the technique to your supplied different project and save with Ctrl+S. Return to the player after editing. Revisit near02:05 to check the revised brief and the instructor's revised operation, then return to your project. Playback, requirement changes and the final deadline continue while the player is hidden. No background tutorial contact sheets, hidden frames, pause/seek controls or automatic window switching are supplied. Native content scores and observed switching/full-screen compliance are reported separately.",
          "video": "evidence/godot/600/gpt/model.mp4",
          "poster": "evidence/godot/600/gpt/poster.png",
          "result": "evidence/godot/600/gpt/result.json",
          "audit": "evidence/godot/600/gpt/audit.json",
          "encoding": "evidence/godot/600/gpt/model.web_encoding.json",
          "sha256": "5ab4921a511a23db1ca2f0fd0d392116b89acef5ba7f97a7c46977146104eca6",
          "bytes": 3372300,
          "partial_scores": {
            "workshop": 0.1,
            "final": 1.0
          },
          "regrade_note": null,
          "media_note": "Actual continuous native X11 screen recording captured at 8 fps, with a 3 fps web encode retaining full elapsed time. Original recordings remain archived; web encoding SHA256 is available beside the MP4. Both applications are full screen; the player disappears while the editor is in front. The inherited scroll executor has an undeclared ±12 bound; GPT600s had ten rejected scroll calls. Scores are diagnostic, not a clean latency ablation.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/godot/600/gpt/model.mp4",
          "download": "../downloads/rcua_learn_godot_002_en.zip",
          "source_directory": "tasks/rcua_learn_godot_002"
        },
        {
          "id": "1",
          "label": "Haiku 5.5 · 600s",
          "model": "Haiku 5.5",
          "score": 0.0,
          "version": "0.2.1-fullscreen-budget600-en.1",
          "model_run_version": "0.2.1-fullscreen-budget600-en.1",
          "evaluation_version": "0.2.1-fullscreen-budget600-en.1",
          "budget_s": 600,
          "instruction": "Work alongside the continuous native-software demonstration. The instructor works on a different example; apply the method to your own open project. You may edit while watching. Playback runs for 180 seconds without pausing or seeking. The final deadline is 420 seconds after playback ends, at 10:00. Save a first draft before the 02:00 workshop collection. That immutable draft is worth 20%; final saved content is worth 80%. Later work cannot recover a missed workshop draft. Partial credit reflects the published components of the actual native project.\n\nUse screenshots and ordinary mouse/keyboard only. No terminals, application scripting consoles, scripts, macros, developer tools, DOM, HTTP calls or editing project source as text. Normal numeric fields, object properties, text labels and native application commands are allowed. The task page has optional technique help and saved-work reviews from a scripted instructor. Help and inference consume wall time. The instructor is simulated, not a human. Instructions and feedback refer to released requirements; a later revision overrides earlier advice. Saving or claiming completion on the task page does not edit or save the native application.\n\nSave the editable project in its existing workspace file. Captures use trusted observation time, not file timestamps. The last captured native file before each collection is used; missing or corrupt final delivery earns zero overall. Runtime or reference failures are reported separately. No symbolic links; maximum file size 64 MiB. Equivalent editable constructions and object renaming are accepted unless a reference label identifies a circuit component. Do not replace the project with a screenshot.\n\nDeliver Harbor.tscn in the supplied project, preserving the supplied read-only project.godot. Each collection scores: two correctly positioned visible bridges 15% each; two matching real box collisions 15% each; continuous collision-supported connection between the fixed platforms 20%; unchanged fixed platforms 10%; current lane/clearance compliance with exactly two bridge pieces 10%. Collision credit is conditional on the corresponding bridge reaching its requested position. Position and dimension tolerance 0.05 m; collision/visual mismatch tolerance 0.02 m. Keep exactly the four supplied floor bodies with one BoxMesh surface and one direct BoxShape3D collision each; do not add floor geometry. Preserve their collision and render layers. Keep the supplied BoxMesh and BoxShape3D primitives; node names may change. No custom scripts, external scene assets, executable resources or replacement flattened images; keep a native editable Node3D scene with ordinary MeshInstance3D and CollisionShape3D nodes.\n\nFULL-SCREEN WINDOW-SWITCHING PILOT: The Lesson Player and Godot editor are two independent FULL-SCREEN applications. Only the front application is visible. When Godot is in front, the complete player, its lesson notes and its video are hidden. There is NO persistent sidebar or overlay. Use Alt+Tab to switch between the real editor and the player. Clicking controls pictured inside the recorded video never edits your project. Keep both applications full screen.\nWatch the operation video, switch to the real editor, apply the technique to your supplied different project and save with Ctrl+S. Return to the player after editing. Revisit near02:05 to check the revised brief and the instructor's revised operation, then return to your project. Playback, requirement changes and the final deadline continue while the player is hidden. No background tutorial contact sheets, hidden frames, pause/seek controls or automatic window switching are supplied. Native content scores and observed switching/full-screen compliance are reported separately.",
          "video": "evidence/godot/600/haiku/model.mp4",
          "poster": "evidence/godot/600/haiku/poster.png",
          "result": "evidence/godot/600/haiku/result.json",
          "audit": "evidence/godot/600/haiku/audit.json",
          "encoding": "evidence/godot/600/haiku/model.web_encoding.json",
          "sha256": "c6701773512f55b840e77566ead7fdfed764c67b3b1ea9501b24826debfcc759",
          "bytes": 871913,
          "partial_scores": {
            "workshop": 0,
            "final": 0
          },
          "regrade_note": null,
          "media_note": "Actual continuous native X11 screen recording captured at 8 fps, with a 3 fps web encode retaining full elapsed time. Original recordings remain archived; web encoding SHA256 is available beside the MP4. Both applications are full screen; the player disappears while the editor is in front. The inherited scroll executor has an undeclared ±12 bound; GPT600s had ten rejected scroll calls. Scores are diagnostic, not a clean latency ablation.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/godot/600/haiku/model.mp4",
          "download": "../downloads/rcua_learn_godot_002_en.zip",
          "source_directory": "tasks/rcua_learn_godot_002"
        },
        {
          "id": "2",
          "label": "GPT‑6.1 Sol · 300s",
          "model": "GPT‑6.1 Sol",
          "score": 0.08,
          "version": "0.2.0-fullscreen-en.1",
          "model_run_version": "0.2.0-fullscreen-en.1",
          "evaluation_version": "0.2.0-fullscreen-en.1",
          "budget_s": 300,
          "instruction": "Work alongside the continuous native-software demonstration. The instructor works on a different example; apply the method to your own open project. You may edit while watching. Playback runs for 180 seconds without pausing or seeking. The final deadline is 120 seconds after playback ends, at 05:00. Save a first draft before the 02:00 workshop collection. That immutable draft is worth 20%; final saved content is worth 80%. Later work cannot recover a missed workshop draft. Partial credit reflects the published components of the actual native project.\n\nUse screenshots and ordinary mouse/keyboard only. No terminals, application scripting consoles, scripts, macros, developer tools, DOM, HTTP calls or editing project source as text. Normal numeric fields, object properties, text labels and native application commands are allowed. The task page has optional technique help and saved-work reviews from a scripted instructor. Help and inference consume wall time. The instructor is simulated, not a human. Instructions and feedback refer to released requirements; a later revision overrides earlier advice. Saving or claiming completion on the task page does not edit or save the native application.\n\nSave the editable project in its existing workspace file. Captures use trusted observation time, not file timestamps. The last captured native file before each collection is used; missing or corrupt final delivery earns zero overall. Runtime or reference failures are reported separately. No symbolic links; maximum file size 64 MiB. Equivalent editable constructions and object renaming are accepted unless a reference label identifies a circuit component. Do not replace the project with a screenshot.\n\nDeliver Harbor.tscn in the supplied project, preserving the supplied read-only project.godot. Each collection scores: two correctly positioned visible bridges 15% each; two matching real box collisions 15% each; continuous collision-supported connection between the fixed platforms 20%; unchanged fixed platforms 10%; current lane/clearance compliance with exactly two bridge pieces 10%. Collision credit is conditional on the corresponding bridge reaching its requested position. Position and dimension tolerance 0.05 m; collision/visual mismatch tolerance 0.02 m. Keep exactly the four supplied floor bodies with one BoxMesh surface and one direct BoxShape3D collision each; do not add floor geometry. Preserve their collision and render layers. Keep the supplied BoxMesh and BoxShape3D primitives; node names may change. No custom scripts, external scene assets, executable resources or replacement flattened images; keep a native editable Node3D scene with ordinary MeshInstance3D and CollisionShape3D nodes.\n\nFULL-SCREEN WINDOW-SWITCHING PILOT: The Lesson Player and Godot editor are two independent FULL-SCREEN applications. Only the front application is visible. When Godot is in front, the complete player, its lesson notes and its video are hidden. There is NO persistent sidebar or overlay. Use Alt+Tab to switch between the real editor and the player. Clicking controls pictured inside the recorded video never edits your project. Keep both applications full screen.\nWatch the operation video, switch to the real editor, apply the technique to your supplied different project and save with Ctrl+S. Return to the player after editing. Revisit near02:05 to check the revised brief and the instructor's revised operation, then return to your project. Playback, requirement changes and the final deadline continue while the player is hidden. No background tutorial contact sheets, hidden frames, pause/seek controls or automatic window switching are supplied. Native content scores and observed switching/full-screen compliance are reported separately.",
          "video": "evidence/godot/300/gpt/model.mp4",
          "poster": "evidence/godot/300/gpt/poster.png",
          "result": "evidence/godot/300/gpt/result.json",
          "audit": "evidence/godot/300/gpt/audit.json",
          "encoding": "evidence/godot/300/gpt/model.web_encoding.json",
          "sha256": "1f73e517d8f5835a6226933e79bd42327446ea5dcd7fa96b71041347641749f5",
          "bytes": 1651709,
          "partial_scores": {
            "workshop": 0.0,
            "final": 0.1
          },
          "regrade_note": null,
          "media_note": "Actual continuous native X11 screen recording captured at 8 fps, with a 3 fps web encode retaining full elapsed time. Original recordings remain archived; web encoding SHA256 is available beside the MP4. Both applications are full screen; the player disappears while the editor is in front. The inherited scroll executor has an undeclared ±12 bound; GPT600s had ten rejected scroll calls. Scores are diagnostic, not a clean latency ablation.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/godot/300/gpt/model.mp4",
          "download": "../downloads/rcua_learn_godot_002_budget300.tar.xz",
          "source_directory": "tasks/rcua_learn_godot_002_budget300"
        },
        {
          "id": "3",
          "label": "Haiku 5.5 · 300s",
          "model": "Haiku 5.5",
          "score": 0.0,
          "version": "0.2.0-fullscreen-en.1",
          "model_run_version": "0.2.0-fullscreen-en.1",
          "evaluation_version": "0.2.0-fullscreen-en.1",
          "budget_s": 300,
          "instruction": "Work alongside the continuous native-software demonstration. The instructor works on a different example; apply the method to your own open project. You may edit while watching. Playback runs for 180 seconds without pausing or seeking. The final deadline is 120 seconds after playback ends, at 05:00. Save a first draft before the 02:00 workshop collection. That immutable draft is worth 20%; final saved content is worth 80%. Later work cannot recover a missed workshop draft. Partial credit reflects the published components of the actual native project.\n\nUse screenshots and ordinary mouse/keyboard only. No terminals, application scripting consoles, scripts, macros, developer tools, DOM, HTTP calls or editing project source as text. Normal numeric fields, object properties, text labels and native application commands are allowed. The task page has optional technique help and saved-work reviews from a scripted instructor. Help and inference consume wall time. The instructor is simulated, not a human. Instructions and feedback refer to released requirements; a later revision overrides earlier advice. Saving or claiming completion on the task page does not edit or save the native application.\n\nSave the editable project in its existing workspace file. Captures use trusted observation time, not file timestamps. The last captured native file before each collection is used; missing or corrupt final delivery earns zero overall. Runtime or reference failures are reported separately. No symbolic links; maximum file size 64 MiB. Equivalent editable constructions and object renaming are accepted unless a reference label identifies a circuit component. Do not replace the project with a screenshot.\n\nDeliver Harbor.tscn in the supplied project, preserving the supplied read-only project.godot. Each collection scores: two correctly positioned visible bridges 15% each; two matching real box collisions 15% each; continuous collision-supported connection between the fixed platforms 20%; unchanged fixed platforms 10%; current lane/clearance compliance with exactly two bridge pieces 10%. Collision credit is conditional on the corresponding bridge reaching its requested position. Position and dimension tolerance 0.05 m; collision/visual mismatch tolerance 0.02 m. Keep exactly the four supplied floor bodies with one BoxMesh surface and one direct BoxShape3D collision each; do not add floor geometry. Preserve their collision and render layers. Keep the supplied BoxMesh and BoxShape3D primitives; node names may change. No custom scripts, external scene assets, executable resources or replacement flattened images; keep a native editable Node3D scene with ordinary MeshInstance3D and CollisionShape3D nodes.\n\nFULL-SCREEN WINDOW-SWITCHING PILOT: The Lesson Player and Godot editor are two independent FULL-SCREEN applications. Only the front application is visible. When Godot is in front, the complete player, its lesson notes and its video are hidden. There is NO persistent sidebar or overlay. Use Alt+Tab to switch between the real editor and the player. Clicking controls pictured inside the recorded video never edits your project. Keep both applications full screen.\nWatch the operation video, switch to the real editor, apply the technique to your supplied different project and save with Ctrl+S. Return to the player after editing. Revisit near02:05 to check the revised brief and the instructor's revised operation, then return to your project. Playback, requirement changes and the final deadline continue while the player is hidden. No background tutorial contact sheets, hidden frames, pause/seek controls or automatic window switching are supplied. Native content scores and observed switching/full-screen compliance are reported separately.",
          "video": "evidence/godot/300/haiku/model.mp4",
          "poster": "evidence/godot/300/haiku/poster.png",
          "result": "evidence/godot/300/haiku/result.json",
          "audit": "evidence/godot/300/haiku/audit.json",
          "encoding": "evidence/godot/300/haiku/model.web_encoding.json",
          "sha256": "a3a3d3fb4dea549a3dfc4bbf9f6e3b5654156df13b338545e0dc29a2cc0caab5",
          "bytes": 473079,
          "partial_scores": {
            "workshop": 0,
            "final": 0
          },
          "regrade_note": null,
          "media_note": "Actual continuous native X11 screen recording captured at 8 fps, with a 3 fps web encode retaining full elapsed time. Original recordings remain archived; web encoding SHA256 is available beside the MP4. Both applications are full screen; the player disappears while the editor is in front. The inherited scroll executor has an undeclared ±12 bound; GPT600s had ten rejected scroll calls. Scores are diagnostic, not a clean latency ablation.",
          "source_url": "https://realtime-cua-benchmark.rongjie.chatgpt.site/video-learning/godot/300/haiku/model.mp4",
          "download": "../downloads/rcua_learn_godot_002_budget300.tar.xz",
          "source_directory": "tasks/rcua_learn_godot_002_budget300"
        }
      ]
    }
  ],
  "evidence_hashes": {
    "evidence/inkscape/teacher_poster.jpg": "a69a8cb236a1a7455f30e08c1fdfa471323222e35f3cd9be2deb6068b138bfc0",
    "evidence/inkscape/tutorial.mp4": "d6fa8b93662a5bd655946f00f903bf8620492290a042ee618f9d206e61200351",
    "evidence/kicad/teacher_poster.jpg": "dc8dcfed99eac1185bbb1a686b82c49bd55d148663fb181f6db37e0ed8cd3345",
    "evidence/kicad/tutorial.mp4": "031e18c22619a6ef6026ffd39f3a0810052d08688c9c249db5326c3bd4e3f7c7",
    "evidence/calc/teacher_poster.jpg": "5c06d9825f912232b49e9dcece36e067e5df0a2f81725de5bf9ad0557a4502f8",
    "evidence/video-self-notes/teacher_poster.webp": "4bcc9ed95e25899865e76b5d372ea6a9220ea3f0152492e9563a3f8caed33124",
    "evidence/impress/teacher_poster.jpg": "0a17d759488bedcd43db09e3e735b09abe3beb842abf667c136ab3ac8e03369f",
    "evidence/impress/tutorial.mp4": "ba3ef5ab49941cf9f538dfbd1a4e063b61f7cb4005a3d430c28d1f527e969c7a",
    "evidence/blender/teacher_poster.jpg": "ddfe254c3f273cbd8fe4a4a0bf2f37317604379368708030fccf3f327b08be1c",
    "evidence/blender/tutorial.mp4": "9c39daca9bcdad5ef3700e7664b13a151fb2a2eaa6afe3e7c50724355cbba0c8",
    "evidence/godot/teacher_poster.jpg": "f18abb73099c3c938f41d13fe975bbf618bbaf3523e2a9aa28c7a3e4ba642b4f",
    "evidence/godot/tutorial.mp4": "4a402e2196ae3350be8ade976e9678dbc90a9d337f717fa77081a05387ef04b4",
    "evidence/video-only/tutorial.mp4": "0ea4b78d32be191cdfa9d3addcc9bf0a7f28e3cf0fcc192738f00029394d6a79",
    "evidence/godot/600/gpt/audit.json": "3531752d6fc555cdb09bec5ce34cc013d1ebd158cbff4f0a302a854334dc0323",
    "evidence/godot/600/gpt/result.json": "f338a518f01d5d00a5127aa1a9bd4e0533dd4f0743f7b38268ab3288d47b281a",
    "evidence/godot/600/gpt/model.web_encoding.json": "eb382857a975180753da2b52ca855f999a4f55a885c056e56ab91765d9ab49ca",
    "evidence/godot/600/gpt/poster.png": "88973d2a97a4be00df561eeee36318d1726556d5ea994a5fc5ae90ace5b2fc60",
    "evidence/godot/600/gpt/model.mp4": "5ab4921a511a23db1ca2f0fd0d392116b89acef5ba7f97a7c46977146104eca6",
    "evidence/godot/600/haiku/audit.json": "fc86d4ae5c7c77136a9fd4c0b8ef614bdef35be17e2cd104ab7ea5b2ef5de9c4",
    "evidence/godot/600/haiku/result.json": "d32fddbf7be93222dc169c3512120bddaf4b550fa22b5fa4204e0ac46b3f1063",
    "evidence/godot/600/haiku/model.web_encoding.json": "430cb3e1f9515734aacc0b84fa74cd2932b1845c85700e368f243dcc6f74268b",
    "evidence/godot/600/haiku/poster.png": "88973d2a97a4be00df561eeee36318d1726556d5ea994a5fc5ae90ace5b2fc60",
    "evidence/godot/600/haiku/model.mp4": "c6701773512f55b840e77566ead7fdfed764c67b3b1ea9501b24826debfcc759",
    "evidence/godot/300/gpt/audit.json": "5c248b23374bf6bbc8c06857758de41323c1c7bddb6521f846dffa38e8446f60",
    "evidence/godot/300/gpt/result.json": "ae6e2dd45f21a77b7a2430d1d1e51fecbba6754121664cbd68935c905d45b337",
    "evidence/godot/300/gpt/model.web_encoding.json": "3dd1deea41c28fd0578ae630302e56e0d637245020f8003b8d3fe6338f7ac77c",
    "evidence/godot/300/gpt/poster.png": "c1f224ccfa6c0bfaa635eec68290935ad66883f61d77dbecebff82f8a43281ac",
    "evidence/godot/300/gpt/model.mp4": "1f73e517d8f5835a6226933e79bd42327446ea5dcd7fa96b71041347641749f5",
    "evidence/godot/300/haiku/audit.json": "c688cc5643d46c4c7686a83fd36f8fd6b58079feb70cb8395243204888283293",
    "evidence/godot/300/haiku/result.json": "e067406d7e8d849e389aec28b5f00e7edbcf601ca43747cc09cc4d15fe614386",
    "evidence/godot/300/haiku/model.web_encoding.json": "e59c22ad0900a1774ff02c169fb5b87a54360f36d45dc65198e9b37e2eeeb304",
    "evidence/godot/300/haiku/poster.png": "9429ea1d7a24a9bb8578bf31a2a5932b94f6895d89aa1e07b525d2d60e105ea8",
    "evidence/godot/300/haiku/model.mp4": "a3a3d3fb4dea549a3dfc4bbf9f6e3b5654156df13b338545e0dc29a2cc0caab5",
    "evidence/video-self-notes/gpt/poster.webp": "d1eaa707f4bb508b21dea84ef4c53cdb1986e4c89b9ab11ad3cec61598721408",
    "evidence/video-self-notes/gpt/encoding.json": "bfa9806810d8688b58fb4124919c8ef612a39454a202d3eb8b3596d9739036d2",
    "evidence/video-self-notes/gpt/result.json": "839210a90be354e4a5c8f0bc7d11826645ff3a95e69f8fb2f6707bb89276744f",
    "evidence/video-self-notes/gpt/model.mp4": "8009661edf2d90464eb3a0349f5c695229d63c407b6ee0be898ea1105494595c",
    "evidence/video-self-notes/gpt/publication_audit.json": "c8d48e36f3632f46493c59630256c85add19b93f43d78fb4421cce5aed3ad685",
    "evidence/video-self-notes/haiku/poster.webp": "fa5528922e2395b3ab66fce13856c9fc665d74dab9ab90d1d20cbc1c22744761",
    "evidence/video-self-notes/haiku/encoding.json": "7d6740eef3f4020a3998b48ef8d60fc870d3b51cf6624b5ee8dc2f4d43bfaca7",
    "evidence/video-self-notes/haiku/result.json": "65832ea84a8c080e0e6fb5a9bc17af20e667e50924e3978ad807a65b2686c536",
    "evidence/video-self-notes/haiku/model.mp4": "38c91841f3a6e8da5c3acc01c4403ae8304f26dfb7e7b40568efee0131c9bca9",
    "evidence/video-self-notes/haiku/publication_audit.json": "afa20c092590ec3113137f28b156a9579eaeac920cc87c974e47d144e84f5e76",
    "evidence/local/blender/gpt/result.json": "7863e43b1707511361b1e90f420273cca62a7586fa9cdacfeb266917f36e57ae",
    "evidence/local/blender/gpt/corrected_evaluation.json": "5e0ed22cdba83fc43822073cf35973f531203ecadcf2e71667222a5cbe7d2eae",
    "evidence/local/blender/gpt/poster.png": "b7daeaa76732e66c3aa45a4e8111cb4d2a9acad7289cbe0f7ec2d2b05b6665d5",
    "evidence/local/blender/gpt/model.encoding.json": "0562d1c2df6f772227484c9d1316ea17ee826f8e024710c4a7431ab5d1219721",
    "evidence/local/blender/gpt/model.mp4": "88371e5aa1ca02a6edbc2c624667de3b57fd6703a4c4815cb33777d8898856cf",
    "evidence/local/blender/gpt/publication_audit.json": "abd78b4d08c306dd57fc904eedcfaddab591c5464ddfb9f3944be05f6f57d76e",
    "evidence/local/blender/haiku/result.json": "1564baef92d227884a7d62a88a6291ded1dacb7e21776767a29830d92d6c47f9",
    "evidence/local/blender/haiku/corrected_evaluation.json": "e6b8c6e5b7dd5487848e7f542a06cb5f955104057b4e7f7e11e4c18707bb5b5a",
    "evidence/local/blender/haiku/poster.png": "04c4794248c2d8371ad66e122db5a39bab9cb776ff93c93b3d9a3d7fe97a0804",
    "evidence/local/blender/haiku/model.encoding.json": "8e92972e1803d59daceade9417a5adeec329e49ce7011c93e0075b5e1f70d7b7",
    "evidence/local/blender/haiku/model.mp4": "9fa40c68afc2670fa108743c50d3705428421b493c3cd2a39ea0abbd8bb1d3d7",
    "evidence/local/blender/haiku/publication_audit.json": "efef21b6abff17d80f0e278b4a44ae7a31bc68ba91ce9d08d54c4f00a04b60fe",
    "evidence/local/impress/gpt/result.json": "72bba5dd41568983d9c3ee2921a527b8b2fa8dbc58413263e35ea02afe150671",
    "evidence/local/impress/gpt/corrected_evaluation.json": "4136dfaa8ef89665010b6e99ece99715d88d5ebdd15c00c9d1165d05c7794f8d",
    "evidence/local/impress/gpt/poster.png": "f76cf3f5ad6ae171a594d133016da4518ebb1e12327a9872aa3a23fb4b02938b",
    "evidence/local/impress/gpt/model.encoding.json": "f0417f1e8372c90b380c31e967b23af5be3813d375ead31078356183510e4a87",
    "evidence/local/impress/gpt/model.mp4": "9515835e0c74edbcdfe367a55e56677e8d0b7da62a43a0a0662a14b29abf60c1",
    "evidence/local/impress/gpt/publication_audit.json": "cec5ad112737df5a4aadc2c82fe81ad58a3075f424ce382c59322eda7291f95c",
    "evidence/local/impress/haiku/result.json": "bb3a13f08407d78379e325c714241866cf949a1f0a280015fa49938335c96432",
    "evidence/local/impress/haiku/corrected_evaluation.json": "2513093ddb8f641b3a6e67d72d9c546def949d783541ce620a1aa04658dac09c",
    "evidence/local/impress/haiku/poster.png": "a090ed13c550153f699a93d2c3a62643d58a2a7038f25f0565b46f5795a0eb9c",
    "evidence/local/impress/haiku/model.encoding.json": "9910054f11fe4f01b68593945f18359f3a1da615252f0ebc59256527dd29e967",
    "evidence/local/impress/haiku/model.mp4": "370e56254597b8fa304a78395aa5233caddcdd400a5c61d88ee5026361fd9c5a",
    "evidence/local/impress/haiku/publication_audit.json": "3398e8c9ecaf9b612f227ecd4152d2397c350a9af25cc367b698856d215a82d0",
    "evidence/local/kicad/gpt/result.json": "7484c33fef40b132501a69f22ab43853a254c6a87c34b641eba48476f63b0cd1",
    "evidence/local/kicad/gpt/corrected_evaluation.json": "b05fe0c929c6926ebac457fc2cdde38178120a8c28ae89e54ae6c88c9dc12486",
    "evidence/local/kicad/gpt/poster.png": "4103c9d13b7800db5d7b89604d9edd767a712fb30d8905ddab1e6653f0b0b74c",
    "evidence/local/kicad/gpt/model.encoding.json": "2501a4d19fa63cb466c4eddce7b86b8c6bf3c7d60efb5f289c61331d3f495b9c",
    "evidence/local/kicad/gpt/model.mp4": "4ed54a24d7691ac990f6c6613e41aebff6dc519006f20b51ba263291ab3598f3",
    "evidence/local/kicad/gpt/publication_audit.json": "2e113eeaf423202ba0be625c7f45226dae9bbe4d7f65715c6c2e61408b067545",
    "evidence/local/kicad/haiku/result.json": "e320aabd8a54b8822a396879eb93594da63f58ef39e668ebda81782f5b06eb61",
    "evidence/local/kicad/haiku/corrected_evaluation.json": "7d0f4d0d5ad043c807a827efde292ce37801a66a76f6721952df6daec841aa43",
    "evidence/local/kicad/haiku/poster.png": "3af3c41bc079c410feed0c7333604641c0dfc1189b83dda751b372adf8cc93ca",
    "evidence/local/kicad/haiku/model.encoding.json": "6dd917fc5656e4dd5e72fdef3bab32c5a71e2549a1631782f65823446dd5b15f",
    "evidence/local/kicad/haiku/model.mp4": "54e5e4441fb2689e005ecd087365d825d7c011d23fcd03fcfda2d403867b2573",
    "evidence/local/kicad/haiku/publication_audit.json": "ef9d0d52c2b087c3a2cc7e07337bf8e0f5d10465c6846e2a18ab9398847bb790",
    "evidence/local/inkscape/gpt/result.json": "2ca67618438c4874a3e4a1d0552e647c5c5b7be634b2c03d2e28b1a358c7f2ee",
    "evidence/local/inkscape/gpt/corrected_evaluation.json": "8d2ba4822cbd0cb08840f291b88640feedda3ad4b4a7598152057d1cf0ac3816",
    "evidence/local/inkscape/gpt/poster.png": "0e244fc91c42d066921aa4b5c10ea3babe14d86fd023b164451d61a722958a93",
    "evidence/local/inkscape/gpt/model.encoding.json": "cbdc090fb178c45fd55b3ec1a2031619032d8a2c349737ebbe3307e998bdef8b",
    "evidence/local/inkscape/gpt/model.mp4": "e1760996e0a6bb6f5cbbd4374645ba9ed6f5968bfc3c043decd17d4f626e2ece",
    "evidence/local/inkscape/gpt/publication_audit.json": "7fd7a4aeee309e64058d6866f8c8c2aa0b87c9bef045d0b34781f98744ad1770",
    "evidence/local/inkscape/haiku/result.json": "33619db0371c005ff0179f5e3cf22bbf128a9a9374ca18241ae5ef667fb6866e",
    "evidence/local/inkscape/haiku/corrected_evaluation.json": "2597bb47670d7d81ef7e4060727757824026c0d94d378ab098ef8e9aa4410863",
    "evidence/local/inkscape/haiku/poster.png": "e9bac7e33989c583133862bbc0b55cbcec7191ec33c2ee668f24cc76ff0fdfe7",
    "evidence/local/inkscape/haiku/model.encoding.json": "2c6ad8eecf3f53612a94ae0e4620b3f1098c7235ac90f2c2fad28780808115a2",
    "evidence/local/inkscape/haiku/model.mp4": "49701f7bca8318415afee07adeae32c0223633bad0f069109e91bda084453b8c",
    "evidence/local/inkscape/haiku/publication_audit.json": "f615282699436fa1d550f6052fe3ab1e6a452b5bd3a1ed9e5c15a559dc4669ca"
  },
  "notes": [
    "Six distinct English native tasks; current complete sources live under tasks/.",
    "Model-run version and evaluator version are different when saved artifacts were regraded. Regrading is not another model execution.",
    "Each listed trial is one local run; native video budgets and player restrictions vary by task. These scores are not pooled with the 17-task leaderboard.",
    "Native apps and the host dependencies must be installed separately. The runtime is not a validated native OSWorld VM integration."
  ]
}