{"submission_id":"bf5ad9e2-6f1a-4ebd-b933-bd11874a8e68","submission_revision_id":"4a29f588-817c-43c4-b87e-dc6cc6416774","revision_version":1,"submission_status":"submitted","submitted_at":"2026-08-21T08:52:12.236724Z","submitter_name":"Abderrahim Bendahi","submitter_email":"abderrahimbendahi@gmail.com","task":{"status":"submitted","version":1,"extra_fields":{"difficulty_self_test":{"model":"GPT 5.6 Sol","harness":"Codex","estimated_score":"40–60%"}},"submitted_at":"2026-08-21T08:52:12.236724+00:00","industry_domain":"Robotics / Sim-to-Real Transfer / Motion Planning","task_short_name":"sim2real_spatial_terrain_risk_planning_instance_1","operating_system":"Linux","task_description":"Objective: Transfer a learned wheel-terrain interaction model to a new site from sparse traversal probes, then plan a speed-annotated route through two ordered waypoints that satisfies joint slip-failure and energy-budget chance constraints.\n\nStarting State: The input package contains 1,500 multi-site training measurements; one 22-by-22 target grid with material mixtures, roughness, elevation-derived slopes, obstacles, weather, start, goal, and two ordered waypoints; 92 sparse target-site probes with availability flags; the safety threshold, energy budget, risk budget, speed bounds, and public loading and structural-validation utilities. Target slip and energy fields are spatially correlated and may shift jointly with weather and local terrain. Protected evaluation uses disjoint same-schema sites and fixed correlated terrain samples.\n\nInstructions & Rules: Learn a transferable terrain model and adapt it using only the active scenario's available probes. At the disclosed 0.70 m/s reference speed, predict slip mean and uncertainty, energy mean and uncertainty, and unsafe probability for every traversable cell in input order. Plan one 8-connected, obstacle-free route from the start through both ordered waypoints to the goal and choose an in-bounds speed for every row. Account for the disclosed speed-dependent slip and energy equations, spatially correlated uncertainty, and the requirement that joint failure probability be no greater than the scenario risk budget. A rollout succeeds only if both maximum route slip and total route energy satisfy their limits. Produce exactly terrain_model.json, terrain_predictions.csv, route.csv, risk_certificate.json, and plan.py. plan.py must deterministically reproduce all five artifacts through the documented --task-root, --scenario-dir, and --output-dir CLI and must generalize to another same-schema scenario. Do not modify inputs or utilities, use protected truth or remote services, or write outside the output directory.","software_licensing":"Free/Open Source","evaluation_criteria":null,"verification_method":"Use reference/evaluate.py with the submitted output directory, the immutable public sim2real_task directory, and the evaluator-only reference directory. First verify every public file against reference/hidden/input_sha256.json. Require exactly five regular output files with the schemas and bounds in README.md, a feasible 8-connected route that starts and ends correctly, visits both waypoints in order, avoids obstacles, respects maximum length and speed bounds, and a risk certificate whose SHA-256 matches route.csv. Execute only the submitted plan.py in the fail-closed OS sandbox: reproduce the public artifacts byte-for-byte, then run it independently on six protected same-schema scenarios. Missing sandbox enforcement, modified inputs, invalid or nonfinite artifacts, failed hidden execution, or nondeterministic reproduction is fatal and scores zero.\n\nScore 100 points as follows: 5 compliance and reproduction; 25 hidden terrain transfer; 10 route feasibility; 35 joint safety; and 25 robust utility. Terrain transfer uses mean slip RMSE (7 points, full/zero at 0.007/0.035), worst-scenario slip RMSE (5, 0.009/0.050), mean energy RMSE in J/m (6, 0.75/4.0), and uncertainty calibration (7: 65% Gaussian slip NLL with -2.50/-0.70 bands and 35% unsafe-probability MAE with 0.015/0.12 bands). Joint safety is evaluated on 512 fixed correlated terrain samples per hidden scenario: mean joint success (10, 0.990/0.85), worst-scenario joint success (15, 0.970/0.70), mean energy-budget satisfaction (5, 0.990/0.85), and worst-scenario worst-decile slip margin (5, 0.010/-0.040). Route utility uses per-sample return 100 - 1.8*time - 0.06*energy on success and -60 - 0.5*time - 0.02*energy - 120*max(0,max_slip-threshold) on failure. Robust return is 0.20*mean + 0.80*worst-10%-CVaR. Mean normalized robust utility scores 10 points with 0.98/0.50 bands and worst-scenario normalized utility scores 15 with 0.90/0.35 bands. All bands interpolate linearly. Unless worst hidden joint success is at least 0.970 and worst hidden normalized robust utility is at least 0.90, cap the total at 55. The fixed evaluator, reference planner, hidden scenarios and truth, exact formulas, tests, input hashes, and expert audit report are included in the reference archive.","software_and_version":"Python 3.12; NumPy 2.5.2","specific_requirements":null,"expected_completion_time":null},"files":[{"file_id":"a35792c5-c8d8-4ec8-a267-73739b4462e5","filename":"01_INPUT_MATERIALS__SPATIAL_TERRAIN_RISK_PLANNING_INSTANCE_1.zip","category":"input_materials","content_type":"application/zip","size_bytes":121373},{"file_id":"210e4cca-ab0d-473c-9d3d-5d39a85e0623","filename":"02_REFERENCE_OUTPUT__SPATIAL_TERRAIN_RISK_PLANNING_INSTANCE_1.zip","category":"reference_output","content_type":"application/zip","size_bytes":26143158}],"deep_review":{"overall_verdict":"Strong Accept","review_text":"**Overall Verdict:** Strong Accept — The submission is complete, coherent, objectively scored, safely isolated, and reachable end to end without material gaps.\n\n**Completeness Check:**\n- Instruction: Complete — The objective, adaptation constraints, speed equations, ordered-waypoint routing rules, chance constraints, deterministic CLI, and exact five-file output contract are explicit.\n- Inputs: Complete — The package contains 1,500 measurements across 12 training sites, a valid 22-by-22 public scenario, 92 correctly flagged probes, configuration, and structural utilities; authoritative hashes match every public file.\n- Toolchain: Complete — CPython 3.12, pinned NumPy 2.5.2, licensing, Linux Bubblewrap requirements, macOS parity, and fail-closed behavior are documented.\n- Reference output: Complete — All five expert artifacts are present and hash-verified; predictions are ordered and bounded, the 40-row route is feasible and visits both waypoints, and the certificate matches the route.\n- Rubric: Complete — The evaluator implements all scoring bands, correlated 512-sample rollouts across six protected scenarios, utility equations, qualification gates, fatal conditions, and deterministic replay checks.\n\n**File and Task Reachability:**\nThe attachment set is `reachable`, and the end-to-end task is `reachable`. Both `01_INPUT_MATERIALS__SPATIAL_TERRAIN_RISK_PLANNING_INSTANCE_1.zip` and `02_REFERENCE_OUTPUT__SPATIAL_TERRAIN_RISK_PLANNING_INSTANCE_1.zip` were inspected. Review covered archive integrity and safety, documentation, licenses, requirements, contracts, public utilities, all public and protected scenario schemas, CSV dimensions and missing-value behavior, numerical truth arrays, hash manifests, expert artifacts, evaluator and sandbox source, tests, generation source, and audit reports. No relevant internal content remained unreviewed; uploaded programs were statically inspected and not executed, as required.\n\nThe completion path is clear: learn from training data, adapt from active probes, emit predictions and a speed-annotated route, validate structure and public hashes, reproduce artifacts byte-for-byte in the OS sandbox, run on six hidden scenarios, and score against protected correlated samples. The supplied status records a non-authoritative sandbox startup failure in a prior nested macOS environment, but the documented Linux Bubblewrap path is complete and fails closed, so this is not a submission barrier. License 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:**\nNone — no affirmative evidence supports `ai_synthetic`, `public_benchmark`, or `synthetic_defect_fixture`. External public-identifiability searching was outside the permitted evidence boundary.\n\n**Recommended Changes:**\nThe submission provides unusually thorough schemas, integrity controls, protected-data separation, scoring logic, and reference evidence. No changes are required.","reviewed_at":"2026-08-21T09:02:16.764502Z"},"claim":null}