{"claim":{"id":"d57932a4-6bb1-456a-9444-f86387d577c7","submission_id":"0b8efdef-a05a-4cc3-a46c-406bf3a306be","submission_revision_id":"a4aaf829-e07c-4b23-a0c8-5c97a2bf10aa","builder":"builder-01","status":"active","claimed_at":"2026-10-15T15:09:27.560751Z","delivered_at":null,"ended_at":null,"delivery_note":null,"end_reason":null},"task":{"submission_id":"0b8efdef-a05a-4cc3-a46c-406bf3a306be","submission_revision_id":"a4aaf829-e07c-4b23-a0c8-5c97a2bf10aa","revision_version":1,"submission_status":"under_review","submitted_at":"2026-08-20T21:01:37.067213Z","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":"2026-08-20T21:01:37.067213+00:00","industry_domain":"Robotics / Sim-to-Real Transfer / Safe Adaptive Control","task_short_name":"sim2real_safe_active_identification_instance_1","operating_system":"Linux","task_description":"Objective: Build a stateful online system-identification and adaptive-deployment controller for a partially observed planar robotic pushing surrogate. The controller must actively identify a correlated actuation/contact regime under independent control and camera delays, commit a calibrated physical belief before the target is revealed, and then transfer that belief into safe target-reaching control.\n\nStarting State: The input package contains the normative NumPy simulator, physical parameter bounds, eight public development cases, an authoritative runner, runtime and license documentation, and structural utilities. Hidden evaluation uses 20 disjoint correlated corner cases. Each case varies anisotropic actuation gains, axis-dependent drag, signed cross-coupling, nonlinear breakaway, control delay, camera delay, camera yaw, camera offset, initial state, masks, and target. Observations contain masked robot-frame encoder position and a separately masked delayed camera-frame position; velocity, applied command, true initial state, delays, transform, and physical coefficients are not revealed.\n\nInstructions & Rules: For each case, run a fresh stateful controller process. Choose 18 safe, informative probe actions while the target is absent from the protocol, then return a bounded uncertainty-aware belief over nine continuous parameters and both discrete delays. The protected parent captures and validates the belief before revealing the target. After target reveal, choose 44 deployment actions. Every action must be finite with Euclidean norm at most 1.0; probe energy must not exceed 8.5; absolute world position must remain at most 0.48 and speed at most 0.72. Produce exactly adaptive_controller.py, agent_config.json, public_beliefs.jsonl, public_rollouts.csv, public_summary.json, and reproduce.py. reproduce.py must regenerate the three evidence artifacts byte-for-byte through the documented task-root/output-dir CLI. Do not alter inputs, infer targets from IDs, access protected paths, or access the network. Protected execution uses an OS sandbox; hidden/reference paths and targets are not mounted into the controller namespace before the protocol reveal.","software_licensing":"Free/Open Source","evaluation_criteria":null,"verification_method":"Use the deterministic protected evaluator in the reference attachment on a Linux worker with Bubblewrap 0.8.0 or newer. First enforce the exact six-file contract, schemas, finite and bounded values, SHA-256 integrity of all 12 public files, the stateful protocol, byte-identical controlled reproduction, and an independent authoritative public rerun. Execute every submitted-code path inside a fail-closed OS sandbox: read-only public/submission mounts as required, no hidden/reference/home mount, no network, no writable controller filesystem, and only one writable reproduction output directory. The evaluator parent loads hidden cases in memory and withholds each target from the protocol and controller namespace until after capturing the belief. Any contract, integrity, sandbox, protocol, or reproduction failure scores zero.\n\nThen execute the controller on 20 fixed disjoint hidden correlated cases and recompute all metrics from runner output. Award 5 raw points for artifact integrity and reproduction; 18 for safe, energy-bounded, two-axis, multilevel active probing; 18 for pre-reveal continuous-parameter accuracy/calibration and control/camera-delay probabilities; and 59 for safe adaptive deployment. Deployment includes 35 points for mean normalized target progress in the worst 20% of hidden cases, with zero at 0.80 or below and full credit at 0.985 or above, plus success rate, final distance, safety, and mean return. Apply the disclosed safety qualification after raw scoring: cap the final score at 55 unless hidden deployment safety is exactly 100% and worst-20% normalized progress is at least 0.95. The report preserves raw score, sections, both qualification predicates, and cap status.\n\nThe six fixed expert artifacts are independently SHA-256-pinned. Their trusted-reference logic audit reproduces the artifacts and scores exactly 100/100 with empty notes, safety 1.0, and worst-20% progress 0.996393. protected_execution_status.json transparently records that the enclosing Codex desktop sandbox prevented a nested sandbox-exec run; the authoritative ALE run must use the documented Bubblewrap command and is valid only with details.sandbox.enforced=true.\n","software_and_version":"Python 3.12; NumPy 2.2.6; Bubblewrap 0.8.0","specific_requirements":null,"expected_completion_time":null},"files":[{"file_id":"c9e2c2a2-50a3-4f1a-b2fd-ff9b8a929c3a","filename":"01_INPUT_MATERIALS__SAFE_ACTIVE_IDENTIFICATION.zip","category":"input_materials","content_type":"application/zip","size_bytes":14888},{"file_id":"2f6f4430-04f8-4827-b151-0c810a2a59e2","filename":"02_REFERENCE_OUTPUT__SAFE_ACTIVE_IDENTIFICATION.zip","category":"reference_output","content_type":"application/zip","size_bytes":90481}],"deep_review":{"overall_verdict":"Strong Accept","review_text":"**Overall Verdict:** Strong Accept — The submission is complete, professionally specified, objectively evaluable, and reachable end to end without material defects.\n\n**Completeness Check:**\n\n- Instruction: Complete — The objective, stateful protocol, safety limits, phase ordering, exact six-file contract, schemas, and reproduction requirements are explicit.\n- Inputs: Complete — The normative simulator, parameter bounds, protocol definition, eight public cases, runtime documentation, and licenses are present; all 12 public-file hashes match the protected manifest.\n- Toolchain: Complete — CPython 3.12, NumPy 2.2.6, Linux Bubblewrap requirements, setup commands, licensing, isolation behavior, and fail-closed conditions are documented.\n- Reference output: Complete — All six expert artifacts are present and match their SHA-256 manifest. Their public evidence has the expected schemas, eight beliefs, and 496 rollout records. The trusted logic report records 100/100 while clearly distinguishing itself from an authoritative sandboxed score.\n- Rubric: Complete — The full 100-point allocation, interpolation thresholds, fatal gates, hidden metrics, safety qualification, and 55-point cap are precise and implemented consistently by the evaluator.\n\n**File and Task Reachability:**\n\nAttachment classification: `reachable`. Task classification: `reachable`.\n\nBoth manifest attachments, `01_INPUT_MATERIALS__SAFE_ACTIVE_IDENTIFICATION.zip` and `02_REFERENCE_OUTPUT__SAFE_ACTIVE_IDENTIFICATION.zip`, were inspected. Their declared sizes match, both are readable unencrypted ZIP containers, and all 47 members were inventoried without traversal paths, links, encryption, unsafe entries, or excessive expansion. Inspection covered all input documentation, configuration and case JSON, simulator source, licenses, evaluator and sandbox source, hidden manifests and cases, six expert artifacts, evidence reports, pipeline code, and tests. JSON, JSONL, and CSV structures were parsed or schema-checked; executable content was reviewed statically and not run. No archive content was left unreviewed.\n\nThe solver-visible input and evaluator-only reference material are separated. The completion path is clear: implement the exact six artifacts, generate public evidence with the authoritative runner, reproduce it byte-for-byte, then evaluate 20 hidden cases under a fresh-process Bubblewrap boundary. The submitted protected-status record appropriately reports that the author-host run was unavailable and requires production execution on the documented Linux worker with `details.sandbox.enforced=true`; this is an infrastructure limitation rather than an attachment defect.\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. Simulator-generated data do not establish AI fabrication or a synthetic-defect fixture. Public identifiability was not independently checked because external search is outside the permitted evidence boundary.\n\n**Recommended Changes:**\n\nThe submission provides a coherent protocol, reproducible expert baseline, strong integrity controls, and a detailed fail-closed evaluator. No contributor changes are required. Production scoring should follow the documented Linux Bubblewrap procedure and accept only a report confirming sandbox enforcement.","reviewed_at":"2026-08-20T21:10:27.705184Z"},"claim":{"id":"d57932a4-6bb1-456a-9444-f86387d577c7","submission_id":"0b8efdef-a05a-4cc3-a46c-406bf3a306be","submission_revision_id":"a4aaf829-e07c-4b23-a0c8-5c97a2bf10aa","builder":"builder-01","status":"active","claimed_at":"2026-10-15T15:09:27.560751Z","delivered_at":null,"ended_at":null,"delivery_note":null,"end_reason":null}},"revision_is_current":true}