{"submission_id":"8e77212c-221b-48ab-8771-0fba5ed20eb8","submission_revision_id":"809bc78f-a629-4287-bdac-d610e8e0699e","revision_version":2,"submission_status":"submitted","submitted_at":"2026-08-21T08:00:33.324628Z","submitter_name":"Abderrahim Bendahi","submitter_email":"abderrahimbendahi@gmail.com","task":{"status":"submitted","version":2,"extra_fields":{"difficulty_self_test":{"model":"GPT 5.6 Sol","harness":"Codex","estimated_score":"20–40%"}},"submitted_at":"2026-08-21T08:00:33.324628+00:00","industry_domain":"Robotics / Sim-to-Real Transfer / Sensor Fusion","task_short_name":"sim2real_async_sensor_fusion_transfer_instance_1","operating_system":"Linux","task_description":"Objective: Calibrate a target robotic rig's asynchronous camera, joint encoders, and force sensor, then build a causal multi-rate estimator that transfers across runs with correlated dynamics, timing, and force-electronics shifts. Reconstruct puck state, physical contacts and 2-D impulses, and calibrated uncertainty under long occlusions and identity-ambiguous target/distractor crossings.\n\nStarting State: The input package contains a trusted two-link robot specification, bounded sensor parameters, a known-layout calibration rig, a moving end-effector marker, encoder logs, a no-contact force baseline, five ten-second public runs, strict loaders and structural checks, and runtime/license documentation. Each run supplies 100 Hz encoders, an unsynchronized approximately 29 Hz rolling-shutter camera, 200 Hz force measurements, and 25 Hz queries beginning after a 2.8 s causal identification prelude. A hidden correlated three-dimensional run vector changes mass, drag, spring, encoder transport/latency/drift, force lag/scale/misalignment/bias drift, and an impulse-triggered dynamics switch. Its prior covariance, affine coefficient equations, and bounds are public, but per-run vectors, trajectories, contacts, and sensor truth are not. Camera candidate IDs are reassigned every frame; randomized distractors cross the target; contact-aligned visual occlusions and sensor-only force transients occur.\n\nInstructions & Rules: Fit the bounded shared camera clock, geometry, distortion, rolling readout, encoder biases, and force biases without collapsing timing effects. Estimate each run's three-dimensional latent vector causally from its identification prelude and continue adapting without future samples. Reconstruct position, velocity, physical contact intervals, 2-D impulse, and row covariance at every query. Use only measurements whose corrected host-equivalent acquisition time is no later than the query, and treat a camera frame as atomically available only after its last detection row. Resolve crossings without candidate IDs; propagate through long occlusions using the disclosed run-conditioned dynamics; compensate force lag, scale, misalignment, and bias drift; reject nonphysical transients; and honor the impulse-triggered mode switch. Every row covariance must be symmetric positive definite, have strictly positive determinant sign, have both eigenvalues in [1e-8, 0.25] m^2 inclusive, and have spectral condition number at most 1e4. Produce exactly sensor_model.json, fusion_config.json, public_reconstruction.csv, public_events.csv, uncertainty.json, and reconstruct.py. reconstruct.py must reproduce all six artifacts byte-for-byte through the documented task-root/data-dir/output-dir CLI and execute on protected same-schema runs. Do not modify inputs, access hidden/reference paths, use the network, rely on absolute paths, or write outside the output directory.","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, exact JSON/CSV schemas and ordering, finite bounded values, and symmetric positive-definite row covariances with strictly positive determinant sign, eigenvalues in [1e-8, 0.25] m^2 inclusive, and spectral condition number at most 1e4. Also enforce SHA-256 integrity of all 21 public files, byte-identical public reproduction, successful protected same-schema execution, and a fail-closed OS sandbox for every submitted-code invocation. The sandbox exposes only the isolated reconstruct.py, minimal read-only CPython runtime and pinned packages, read-only public task, read-only selected sequence data, and one writable output directory; it denies the reference/hidden tree, home/workspace, and network. Any contract, integrity, covariance, sandbox, execution, causality, or reproduction failure scores zero.\n\nAward 5 points for compliance and protected reproduction; 15 for shared sensor identification and per-run three-dimensional latent adaptation; 50 for causal reconstruction; 15 for contact intervals, boundary timing, 2-D impulse, and confidence Brier score; and 15 for row-covariance calibration. Reconstruction awards 35 points across public and hidden global error, worst hidden run, long occlusion, randomized target/distractor crossing, tail, and worst-run tail position/velocity RMSE, plus 15 points from eleven full-history-preserving prefix reruns. Prefix data include only complete candidate frames and final prefix estimates must agree with full-log estimates within 6e-4, including covariance, timestamps, and association status.\n\nUncertainty scores 95% Mahalanobis coverage and Gaussian position NLL independently on global, long-occlusion, contact, crossing, and tail strata. Gaussian NLL is computed from each validated covariance by unjittered Cholesky factorization and triangular solve; no pseudoinverse is used. Failed factorization, nonpositive determinant sign, out-of-bound eigenvalue or condition number, empty uncertainty stratum, or nonfinite Mahalanobis, log-determinant, row-NLL, coverage, or mean-NLL value is fatal. Gaussian NLL may legitimately be negative because a continuous density can exceed one. EVALUATION.md publishes every continuous full/zero band and exact formula.\n\nThe six fixed expert artifacts have an independent protected SHA-256 manifest. The hash-gated trusted-reference logic audit reproduces and scores them exactly 100/100 with all 15 uncertainty points, empty notes, and prefix invariance 1.0. 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 when metrics.sandbox.enforced=true.","software_and_version":"Python 3.12; NumPy 2.5.2; SciPy 1.18.0; Bubblewrap 0.8.0","specific_requirements":null,"expected_completion_time":null},"files":[{"file_id":"680d605a-a194-4f5d-b036-1196a0d7cfe3","filename":"01_INPUT_MATERIALS__ASYNC_SENSOR_FUSION__MINOR_REVISION.zip","category":"input_materials","content_type":"application/zip","size_bytes":435359},{"file_id":"13a4fced-7062-49db-a075-34b96369f1b7","filename":"02_REFERENCE_OUTPUT__ASYNC_SENSOR_FUSION__MINOR_REVISION.zip","category":"reference_output","content_type":"application/zip","size_bytes":405544}],"deep_review":{"overall_verdict":"Strong Accept","review_text":"**Overall Verdict:** Strong Accept — The submission is complete, internally coherent, professionally challenging, objectively scorable, and reachable through the documented protected Linux workflow.\n\n**Completeness Check:**\n- Instruction: Complete — The README defines the causal estimation objective, timing and association rules, physical model, exact six-file contract, schemas, bounds, and portable CLI.\n- Inputs: Complete — The input archive contains calibration data, five public sequences, 900 ordered queries, manifests, trusted robot parameters, strict loaders, integrity metadata, and licensing.\n- Toolchain: Complete — CPython 3.12, exact NumPy and SciPy versions, Bubblewrap requirements, setup steps, sandbox behavior, runtime limits, and open-source licensing are documented.\n- Reference output: Complete — All six expert artifacts, four hidden same-schema sequences, 720 hidden queries, truth arrays, SHA-256 manifests, evaluator sources, security tests, and a hash-gated 100/100 logic audit are present and consistent. Native protected execution remains explicitly pending rather than being misrepresented as authoritative.\n- Rubric: Complete — The 100-point allocation, continuous full/zero bands, fatal conditions, covariance formulas, event matching, uncertainty metrics, and eleven prefix checks are precisely specified.\n\n**File and Task Reachability:**\nThe attachment set is **reachable**, and the end-to-end task is **reachable**. Both uploaded ZIP files were inspected: their complete 27- and 37-entry inventories are readable, unencrypted, CRC-valid, and free of traversal paths, links, device entries, unexpected executables, or excessive expansion. Inspection covered all CSV and JSON members, the README, runtime and evaluation documentation, manifests, licenses and notices, loaders, evaluator and sandbox sources, tests, generator sources, the six expert artifacts, integrity manifests, and both truth NPZ inventories. CSV headers, field counts, sequence counts, timing ranges, missing-value conventions, and representative records are consistent; every authoritative public-input and expert-output SHA-256 digest matches. NPZ member names, formats, dtypes, and shapes were inspected, but object-array payloads were not deserialized because doing so would violate safe evidence handling. Native CPython 3.12 and Bubblewrap execution was unavailable in the review environment, so uploaded programs were not run; the supplied non-authoritative logic audit and static sandbox tests were assessed instead.\n\nThe completion path is to use the public calibration and sequence data to produce exactly six artifacts through the documented CLI. The evaluation path verifies public integrity, schemas, bounds, covariance validity, byte-identical public reproduction, protected same-schema execution, hidden truth metrics, event scoring, uncertainty calibration, and eleven sandboxed prefix reruns. The evaluator fails closed unless Bubblewrap reports enforced isolation. 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 — there is no affirmative evidence for an applicable label. Public identifiability was not externally checked because external search is outside the permitted evidence boundary.\n\n**Recommended Changes:**\nThe submission already provides unusually strong contracts, integrity controls, evaluator transparency, protected-data separation, and reference evidence. No contributor changes are required; the documented authoritative Linux Bubblewrap run should occur on the evaluation worker as part of normal evaluation.","reviewed_at":"2026-08-21T08:08:52.290771Z"},"claim":null}