Logic & contraption puzzles
Solve grid and drag-and-place physics puzzles.
20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.
- Typical applications
- Sokoban · Baba Is You · The Incredible Machine
- Class GUI-need label
G3· GUI-only
Task proposals · 20
Expand to inspect the complete original wording and proposed checks. Original proposal text remains English in both interfaces.
R13.02.01.03-T01 Push all boxes onto goals in a fixed Sokoban level
Original instruction
A self-hosted Sokoban web game shows level 3 of the bundled set (4 boxes). Move the worker with the arrow keys so every box ends on a goal square. Undo and restart are allowed.
- Platforms
web· Any OS (web)- Applications
- Sokoban
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
key· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Game-state bridge (harness only): final level array has every box on a goal; level-complete event logged. The level data object is closure-scoped, DevTools are disabled by browser policy, and moves must be trusted (isTrusted) key events, so state cannot be patched by script.
Proposed partial credit
25% one box on a goal at best state; 25% two boxes; 25% three boxes; 25% all four boxes
R13.02.01.03-T02 Reach the 512 tile in 2048 on a fixed seed
Original instruction
A self-hosted 2048 web game (seed 42, standard 4×4) is open. Slide tiles with the arrow keys and create a 512 tile before the board locks.
- Platforms
web· Any OS (web)- Applications
- 2048
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
key· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
JS bridge: max tile value per game; pass at ≥ 512. Game manager object is not exposed on window and DevTools are disabled by browser policy, so the board cannot be patched by script; trusted (isTrusted) key events only.
Proposed partial credit
33% max tile 128 reached; 33% max tile 256; 34% max tile 512
R13.02.01.03-T03 Win a seeded beginner Minesweeper board
Original instruction
A Minesweeper app (Windows: Microsoft Minesweeper offline test profile; macOS: a local Minesweeper app) is on a seeded Beginner board (9×9, 10 mines, first click safe). Clear all safe cells without detonating a mine; flagging is optional.
- Platforms
win· Windowsmac· macOS- Applications
- Minesweeper
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
click· Point & clickvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Game save/statistics: seeded game won. Fallback: final screenshot parsed for revealed grid = all safe cells. The agent has no shell, and the harness logs which process writes the stats/save file, so a win not recorded by the app is detected.
Proposed partial credit
50% scaled by the first half of the 71 safe cells revealed (0 if a mine is hit before then); 50% scaled by the remaining safe cells
R13.02.01.03-T04 Solve a local Wordle in at most five guesses
Original instruction
A self-hosted Wordle clone with a fixed hidden answer and the standard word list is open. Type guesses and use the colour feedback to find the word in 5 guesses or fewer.
- Platforms
web· Any OS (web)- Applications
- Wordle
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 1 / 5
Proposed verifier
Server-side guess log: solved flag and guess count ≤ 5. Answer is never sent to the client (server-side check per guess). The server issues no API token and accepts guesses only from the CSRF-bound browser session, so the log reflects typed guesses.
Proposed partial credit
60% solved within 6 guesses; 40% solved within 5 guesses
R13.02.01.03-T05 Clear the first level of Baba Is You
Original instruction
Baba Is You (desktop) is at level 0 'Baba Is You' of the map. Push the rule blocks or reach the flag to make WIN happen, then also clear level 1 'Where Do I Go?'.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Baba Is You
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
key· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Save file (ba.ba world progress) parsed: levels 0 and 1 marked complete. No level editor use (editor flag absent). The agent has no shell or text editor, and writes to the save folder are logged by process, so a save not written by the game is detected.
Proposed partial credit
40% level 0 complete; 60% level 1 complete
R13.02.01.03-T06 Place parts so the ball lands in the basket in The Incredible Machine
Original instruction
The Incredible Machine (DOSBox, harness-supplied) is on puzzle 1. Drag the provided parts from the parts bin onto the playfield, adjust their positions so the ball falls into the basket, and press Start.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- The Incredible Machine
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
drag· Drag & dropgeom· Precise geometryvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Puzzle-complete screen detected by perceptual hash and RAM puzzle index incremented (DOSBox memory read by harness). Agent input is mouse/keyboard on the DOSBox window only; the DOSBox debugger and config file are not accessible to the agent, so memory cannot be patched.
Proposed partial credit
30% parts placed on playfield; 70% puzzle solved
R13.02.01.03-T07 Solve an Incredible Machine puzzle within ten minutes using only given parts
Original instruction
In The Incredible Machine (DOSBox), open puzzle 3. Solve it within 10 minutes of wall-clock time using no more than the provided parts, with ropes, belts and pulleys connected correctly.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- The Incredible Machine
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
drag· Drag & dropgeom· Precise geometryvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
RAM puzzle index increments within 600 s of task start; completion screen hash. Belt/rope attachment requires precise drag to connection points. The DOSBox debugger and config file are not accessible to the agent, so memory cannot be patched; input is mouse/keyboard on the window only.
Proposed partial credit
30% mechanical links connected; 70% solved within the limit
R13.02.01.03-T08 Bring both characters to their exits in Fireboy and Watergirl
Original instruction
A self-hosted HTML5 build of a Fireboy-and-Watergirl-style co-op puzzle (Forest Temple level 1) is open. Control both characters (arrows and WASD) so each reaches its own exit door without touching the opposite liquid.
- Platforms
web· Any OS (web)- Applications
- Fireboy and Watergirl
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
key· Shortcutsrt· Real-time controlvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Game-state bridge: both characters on exits simultaneously → level-complete event. Deaths logged. Game objects are closure-scoped and DevTools are disabled by browser policy, so state cannot be patched by script; only trusted (isTrusted) keys are accepted.
Proposed partial credit
30% Fireboy reaches his exit; 30% Watergirl reaches hers; 40% level complete
R13.02.01.03-T09 Place two ramps so a ball rolls into a cup
Original instruction
A self-hosted Box2D puzzle page shows a ball on a shelf and a cup across the screen. Drag and rotate two ramp pieces (rotation via handle) into the scene so that, when you press Play, the ball ends inside the cup.
- Platforms
web· Any OS (web)- Applications
- physics puzzle web
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropgeom· Precise geometryvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Deterministic simulation from the saved piece transforms: ball centre inside the cup bounds at rest. Piece transforms accept only pointer interactions; numeric inputs are absent. DevTools are disabled by browser policy and the app issues no API token, so transforms cannot be set by script.
Proposed partial credit
20% two ramps placed; 30% ball reaches the cup side of the screen; 50% ball in cup
R13.02.01.03-T10 Build a bridge in Poly Bridge that holds the truck within budget
Original instruction
Poly Bridge 2 level 1-6 is open. Design a bridge with road, wood and steel members under the level budget so the vehicles cross and reach the flags. Run the simulation to pass.
- Platforms
win· Windowsmac· macOSlin· Linux- Applications
- Poly Bridge
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
drag· Drag & dropgeom· Precise geometryvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Profile save parsed: level 1-6 completed with cost ≤ budget. Sandbox mode and cheats unavailable in campaign. The agent has no shell, and writes to the save folder are logged by process, so a save not written by the game is detected.
Proposed partial credit
30% road spans the gap; 30% within budget; 40% vehicles reach checkpoints
R13.02.01.03-T11 Complete 'Going Up' in World of Goo with extra goo balls
Original instruction
World of Goo, chapter 1 level 'Going Up'. Build a tower of goo balls to reach the pipe and collect at least 15 goo balls (above the level's required 6).
- Platforms
win· Windowsmac· macOSlin· Linux- Applications
- World of Goo
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
drag· Drag & dropgeom· Precise geometryvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Profile file parsed: level GoingUp completed with balls collected ≥ 15. The agent has no shell or text editor, and writes to the profile folder are logged by process, so a profile not written by the game is detected.
Proposed partial credit
40% level completed; 60% scaled by balls from 6 to 15
R13.02.01.03-T12 Earn three stars on a Cut the Rope level
Original instruction
On the Android emulator, Cut the Rope is on level 1-5. Swipe to cut ropes at the right moments so Om Nom eats the candy after it collects all 3 stars.
- Platforms
and· Android- Applications
- Cut the Rope
- Interfaces
mob· Mobile touch- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
touch· Touch gesturespath· Continuous pathrt· Real-time control- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
App save (pulled by adb by the harness) shows level 1-5 with 3 stars; fallback: results screen OCR. adb and root are not exposed to the agent, so the save cannot be edited; input comes only from emulator touches.
Proposed partial credit
40% level completed; 60% scaled by stars (one third per star)
R13.02.01.03-T13 Solve a 15×15 nonogram by filling cells
Original instruction
A self-hosted nonogram (picross) web app shows a 15×15 puzzle with row and column clues and a unique solution. Fill and cross out cells by clicking or click-dragging until the image is solved.
- Platforms
web· Any OS (web)- Applications
- nonogram web
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clickdrag· Drag & dropvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Server stores the grid on every change: final grid equals the unique solution; solved event logged. Clue data is drawn on a canvas only. The server issues no API token and accepts grid changes only from the CSRF-bound browser session, and DevTools are disabled by policy.
Proposed partial credit
50% scaled by (correct fills − wrong fills) up to half of the solution cells; 50% scaled from half to all
R13.02.01.03-T14 Solve an Opus Magnum puzzle under a cost limit
Original instruction
Opus Magnum, puzzle 'Stabilized Water' (chapter 1). Build a machine of arms, glyphs and an instruction tape that produces the required outputs, with total cost ≤ 70 gold, and run it to completion.
- Platforms
win· Windowsmac· macOSlin· Linux- Applications
- Opus Magnum
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
drag· Drag & dropgeom· Precise geometryclick· Point & click- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 5 / 5
Proposed verifier
Solution file (.solution) saved by the game parsed and verified with an open-source Opus Magnum simulator: produces required outputs; cost ≤ 70. The agent has no shell, and the solution file's writing process is logged, so a file not written by the game is detected.
Proposed partial credit
60% valid solution; 40% cost ≤ 70 (scaled from 120)
R13.02.01.03-T15 Solve a Human Resource Machine level within the size target
Original instruction
Human Resource Machine, level 'Zero Exterminator'. Drag command blocks into the program to pass the level's tests, meeting the size challenge (≤ 4 commands).
- Platforms
win· Windowsmac· macOSlin· Linux- Applications
- Human Resource Machine
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
drag· Drag & dropclick· Point & click- Work product
WP06· Code/engineering- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Save file parsed: level solved; best size ≤ 4. Clipboard paste of program text is disabled by blocking the clipboard in the VM.
Proposed partial credit
60% level solved; 40% size target
R13.02.01.03-T16 Solve a hard Sokoban level from the Microban set
Original instruction
The self-hosted Sokoban web game is on Microban level 50 (a harder, multi-box level). Solve it; undo is allowed.
- Platforms
web· Any OS (web)- Applications
- Sokoban
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
key· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Game-state bridge: level complete. Partial credit uses best state boxes-on-goals count. The level data object is closure-scoped and DevTools are disabled by browser policy, so state cannot be patched by script; moves must be trusted (isTrusted) key events.
Proposed partial credit
50% scaled by boxes on goals up to half the boxes; 50% scaled by the remaining boxes (level complete = full)
R13.02.01.03-T17 Clear three levels of Baba Is You World 2
Original instruction
Baba Is You save with the Lake area unlocked. Clear any three levels in the Lake (world 2) by rearranging rules.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Baba Is You
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
key· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
long· Long-horizon- Proposed difficulty
- 5 / 5
Proposed verifier
Save file parsed: number of completed Lake levels increased by ≥ 3. The level editor flag must be absent; the agent has no shell or text editor, and writes to the save folder are logged by process, so a save not written by the game is detected.
Proposed partial credit
33% first level cleared; 33% second level; 34% third level
R13.02.01.03-T18 Win a seeded expert Minesweeper board
Original instruction
A local Minesweeper web build is on a seeded Expert board (30×16, 99 mines) generated as no-guess (always solvable by logic). Clear all safe cells without detonating a mine.
- Platforms
web· Any OS (web)- Applications
- Minesweeper
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clickvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
long· Long-horizon- Proposed difficulty
- 5 / 5
Proposed verifier
Server-side board (mines never sent to the client; reveals computed server-side): game won. Partial credit by safe cells revealed. The server issues no API token and accepts reveals only from the CSRF-bound browser session, and DevTools are disabled by policy.
Proposed partial credit
50% scaled by the first half of the 381 safe cells (0 if a mine is hit before a quarter are revealed); 50% scaled by the remaining safe cells
R13.02.01.03-T19 Assemble a tangram silhouette by dragging and rotating pieces
Original instruction
A self-hosted tangram web app shows a target silhouette (a running man). Drag the 7 pieces, rotate in 45° steps and flip the parallelogram so they fill the silhouette without overlap.
- Platforms
web· Any OS (web)- Applications
- tangram web
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropgeom· Precise geometryvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Final piece transforms from the app: union IoU with the silhouette ≥ 0.98 and pairwise overlap area < 1% of piece area. Transforms are accepted only from pointer drags and rotate/flip buttons; DevTools are disabled by browser policy and no API token is issued, so pieces cannot be placed by script.
Proposed partial credit
50% scaled by silhouette IoU from 0.6 to 0.79; 50% scaled from 0.79 to 0.98; minus one fifth for overlaps
R13.02.01.03-T20 Solve a hard Sudoku using pencil marks
Original instruction
A self-hosted Sudoku web app shows a hard puzzle with a unique solution. Fill all cells; the app supports pencil marks and highlights conflicts only on request (conflict check off). Submit when complete.
- Platforms
web· Any OS (web)- Applications
- Sudoku web
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & formsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Server-side check of the submitted grid against the unique solution. Puzzle givens are rendered to canvas and not present as text in the DOM. The server issues no API token and accepts submissions only from the CSRF-bound browser session, and DevTools are disabled by policy.
Proposed partial credit
80% scaled by correct non-given cells; 20% no wrong cell in the submitted grid
Mapped research references
Existing benchmark-to-class links from the research catalog. They are not record-level annotations, executions or product endorsements.