{"submission_id":"c24aef49-2662-4b49-abb0-d70d984ddb71","submission_revision_id":"237e6283-8ffb-47da-a27c-83c1a387c1bc","revision_version":1,"submission_status":"under_review","submitted_at":"2026-10-10T01:07:01.599248Z","submitter_name":"Ipek Oztas","submitter_email":"ipek_oztas@brown.edu","task":{"status":"submitted","version":1,"extra_fields":{},"submitted_at":"2026-10-10T01:07:01.599248+00:00","industry_domain":"Computer Science (Visual Computing)","task_short_name":"ale_4d_mpm_finite_pulse_collision_001","operating_system":"Linux, MacOS","task_description":"Reconstruct the supplied video as an executable 3D simulation evolving over time.\n\nThe scene contains finite liquid and granular pulses moving toward each other between two opposing pipes inside a stationary glass enclosure. Reproduce their collision, spreading, falling, and subsequent contact with the floor and walls. Motion is passive after initialization; no additional particles are emitted.\n\nStarting materials: reference.mp4, camera.json, scene_facts.json, TASK.md and FORMAT.md. Use the supplied camera calibration and dimensions. Infer the initial conditions, material parameters and detailed dynamics from the video.\n\nWrite a standalone program and render the reconstruction in Blender at 640 × 400 pixels, 32 fps, for all 49 frames. Export the camera, triangle geometry for every frame, and persistent 3D particle trajectories separately for liquid and sand.\n\nDeliver solution/ containing source code, dependency information and an executable build.sh, plus world/ containing the rendered video and exported 4D scene. The build must regenerate the outputs without network access, reference-video access or baked reference trajectories. Copied video textures and flat video planes are prohibited.\n\nThe reference uses MPM-lite; another open-source method is acceptable. Success requires matching the material-specific geometry and motion, passing the numerical thresholds in FORMAT.md, and demonstrating a reproducible build whose exported geometry matches its rendered video.","software_licensing":"Free/Open Source","evaluation_criteria":null,"verification_method":"Run the supplied evaluate.py against the reference and agent world/ directories.\n\n1. Require a decodable 640 × 400 H.264 video containing exactly 49 frames at 32 fps. Require valid camera data, per-frame triangle meshes, and finite, persistent 3D trajectories separately for water and sand. Missing or malformed outputs fail.\n\n2. Require a numerical score of at least 80/100. The evaluator compares material-specific shape (30 points), persistent trajectories (20), displacement distributions (20), centroids (10), spatial spread (10), and occupancy overlap across three orthogonal views (10).\n\n3. Additionally, each material must satisfy all three limits:\n   - Mean symmetric nearest-neighbour shape error ≤ 0.025 m.\n   - Mean assigned trajectory error ≤ 0.06 m.\n   - Sliced displacement-distribution error ≤ 0.005 m per frame.\n\n4. Require the supplied camera calibration: maximum absolute difference ≤ 0.01 for intrinsics and ≤ 0.0001 for the camera-to-world matrix.\n\n5. Rebuild the output from source in a fresh, offline environment without access to the reference video, private ground truth or baked trajectories. Rebuilt geometry and trajectories must match the submitted numerical arrays within an absolute tolerance of 0.00001 m.\n\n6. Audit frames 0, 8, 16, 24, 32 and 48 to confirm that the rendered surfaces match the exported geometry. Reject copied video textures, flat video planes, hidden substitute geometry, omitted visible surfaces, or particle identities reset between frames.\n\nSuccess requires every condition above; a numerical score alone is insufficient.","software_and_version":"MPM-lite 0.1.0, Blender 5.2.1 LTS, Python 3.12.14, NVIDIA Warp 1.10.1 (CPU mode), NumPy 2.3.5, SciPy 1.18.1.","specific_requirements":null,"expected_completion_time":null},"files":[{"file_id":"4f67cb1d-7354-470d-bdfb-e865dd954abc","filename":"TASK.md","category":"input_materials","content_type":"text/markdown","size_bytes":2857},{"file_id":"333e8fdb-b146-49c0-a291-6efac28ab255","filename":"FORMAT.md","category":"input_materials","content_type":"text/markdown","size_bytes":3749},{"file_id":"16dbcb7b-5443-432d-a036-202c98222d35","filename":"scene_facts.json","category":"input_materials","content_type":"application/json","size_bytes":1177},{"file_id":"1995e3be-3594-41db-af3d-8c0c27231360","filename":"camera.json","category":"input_materials","content_type":"application/json","size_bytes":583},{"file_id":"47bfefe3-0508-4cd5-a281-6fc030328078","filename":"reference.mp4","category":"input_materials","content_type":"video/mp4","size_bytes":348119},{"file_id":"5b7ae6ff-d094-490f-b176-6a150c5c5e3f","filename":"render.mp4","category":"reference_output","content_type":"video/mp4","size_bytes":348119},{"file_id":"800ae564-5ea9-4cd8-b455-9b43c3a0ac18","filename":"camera.json","category":"reference_output","content_type":"application/json","size_bytes":583},{"file_id":"04d49506-5722-4206-a34d-916bbd49629f","filename":"evaluate.py","category":"reference_output","content_type":"text/x-python-script","size_bytes":6344},{"file_id":"10c82646-e778-4677-af4c-da2908850632","filename":"requirements-lock.txt","category":"reference_output","content_type":"text/plain","size_bytes":651},{"file_id":"f8685556-ff0f-4c67-aff4-ce7ea766c9cd","filename":"solution.zip","category":"reference_output","content_type":"application/zip","size_bytes":66861},{"file_id":"f12f80f5-1690-442a-85b2-e1d46d4eb2be","filename":"dynamics.zip","category":"reference_output","content_type":"application/zip","size_bytes":1671913},{"file_id":"a9c12958-a00c-4e09-a1ec-9601606a7193","filename":"meshes.zip","category":"reference_output","content_type":"application/zip","size_bytes":51256026}],"deep_review":{"overall_verdict":"Accept","review_text":"**Overall Verdict:** Accept — The submission defines a coherent, professionally substantial reconstruction task with complete numerical references and explicit safeguards against shortcut solutions; no material submission defect was established.\n\n**Completeness Check:**\n\n- Instruction: Complete — Deliverables, timing, calibration, passive dynamics, prohibited shortcuts, and offline regeneration requirements are explicit.\n- Inputs: Complete — The video, camera calibration, scene dimensions, and output specification are supplied and structurally consistent.\n- Toolchain: Complete — Procedural source, vendored simulation engine, dependency versions, executable build script, and an Apache-2.0 license provide a plausible single-machine CPU workflow.\n- Reference output: Complete — Separate material trajectories, ten mesh archives, calibrated camera, and rendered video cover the required artifacts.\n- Rubric: Complete — Weighted metrics, material-specific hard limits, malformed-output rejection, rebuild comparison, and mandatory visual audits define success objectively.\n\n**File and Task Reachability:**\n\nAttachment classification: `partially_reachable`. All 12 manifest attachments were downloaded at their declared sizes and inspected: TASK.md, FORMAT.md, scene_facts.json, both camera.json files, reference.mp4, render.mp4, evaluate.py, requirements-lock.txt, solution.zip, dynamics.zip, and meshes.zip. Archives were safely inventoried and passed internal integrity checks.\n\nBoth materials contain finite float32 trajectories shaped `(49, 2052, 3)` with separate mesh identifiers. Every mesh supplies valid vertices and triangle indices for all 49 frames; eight meshes remain exactly static. Sand geometry centers agree with its trajectories. Camera files are identical and contain a valid pose. Both videos are byte-identical, with container metadata specifying H.264, 640 × 400, 49 frames, and 32 fps.\n\nVideo decoding and visual inspection were unavailable. Evaluator execution, offline regeneration, and rebuilt-array agreement were not verified. Two supplementary documents referenced by the README are absent; their contents remain unreviewed.\n\nTask classification: `reachable`. The supplied inputs support reconstruction, procedural simulation, surface extraction, rendering, and export. Evaluation combines private numerical comparison with isolated rebuilding and calibrated/novel-view geometry audits. Reference source and ground truth are explicitly excluded from solver access.\n\nLicense keys, accounts, tokens, and other credentials are not required from the contributor. Confidential, proprietary, or internal material may be evaluated as submitted evidence without requiring additional disclosure.\n\n**Special Screening:**\n\nNone. No affirmative evidence supports a screening label. Public benchmark provenance and exact software-release availability were not independently verified because approved external search was unavailable.\n\n**Recommended Changes:**\n\nThe calibrated inputs, persistent material references, and explicit manual gates form a strong evaluation package. No required changes were identified. Optionally, make the README self-contained by incorporating its referenced setup and simulation-limitations documentation.","reviewed_at":"2026-10-10T05:51:33.643909Z"},"claim":null}