{"claim":{"id":"4c0e7dfe-6288-4de0-a402-38303a76ee30","submission_id":"53adf6e2-ab55-4030-91ae-803073f1c7b9","submission_revision_id":"68b467da-9afe-4d8d-b6be-5c50c9722c81","builder":"builder-01","status":"active","claimed_at":"2026-10-15T15:04:26.780353Z","delivered_at":null,"ended_at":null,"delivery_note":null,"end_reason":null},"task":{"submission_id":"53adf6e2-ab55-4030-91ae-803073f1c7b9","submission_revision_id":"68b467da-9afe-4d8d-b6be-5c50c9722c81","revision_version":2,"submission_status":"under_review","submitted_at":"2026-08-16T07:35:17.664093Z","submitter_name":"Ziao Yang","submitter_email":"ziaoyang@brandeis.edu","task":{"status":"submitted","version":2,"extra_fields":{},"submitted_at":"2026-08-16T07:35:17.664093+00:00","industry_domain":"AI Engineering / Adversarial Robustness","task_short_name":"ip_ensemble_adversarial_defense_1","operating_system":"Linux","task_description":"Own the scheduled release review for a candidate binary abuse-detection checkpoint trained from six clean-room simulated acquisition programs. Execute ten approved norm-bounded white-box threat scenarios against the fixed dropout ensemble; compute per-record clean-utility and adversarial-robustness inner-product influence statistics; design a budgeted, source-aware trim/relabel/reweight remediation; deterministically retrain; evaluate every release gate; and write attacks.npz, influence_scores.csv, intervention_manifest.csv, predictions.npz, validation.json, release_review.md, plus a reusable defense_pipeline.py. A read-only checkpoint-triage utility supplies ordinary baseline metrics and emits no scored artifact. The evaluator replays the pipeline on an independently provisioned clean-room case.","software_licensing":"Free/Open Source","evaluation_criteria":null,"verification_method":"A deterministic clean-room scorer validates all output schemas, exact record coverage, candidate and freshly regenerated post-remediation attack coordinates and norm budgets, ensemble influence values and stable ranks, source-aware interventions and retention floors, final clean/attacked predictions, every release-gate value, the approve/block decision, and substantive residual-risk analysis. The visible case contributes 35% and an independently seeded private replay contributes 65%. References remain privileged, all candidate paths must be bounded single-link regular files rather than symlinks, and the submitted pipeline runs as an unprivileged account with only whitelisted private inputs. Gold replay outputs score 1.0. Missing or malformed candidate outputs receive component-level zero credit; hidden-reference or evaluator infrastructure failures raise.","software_and_version":"Python 3.12.13, NumPy 2.3.5","specific_requirements":null,"expected_completion_time":null},"files":[{"file_id":"d2c3b7e5-3536-4079-ab28-a92d2d96f0fa","filename":"input_materials.zip","category":"input_materials","content_type":"application/zip","size_bytes":17853},{"file_id":"50bf290f-cd03-49f5-b810-c2919bb1e4d9","filename":"reference_output_and_evaluation_dependencies.zip","category":"reference_output","content_type":"application/zip","size_bytes":72700}],"deep_review":{"overall_verdict":"Strong Accept","review_text":"**Overall Verdict:** Strong Accept — The submission is complete, internally consistent, professionally representative, safely isolated, and objectively evaluable through visible scoring and private deterministic replay.\n\n**Completeness Check:**\n\n- Instruction: Complete — The method specification defines equations, attack procedures, stable ordering, intervention policy, retraining, schemas, release gates, CLI behavior, and required deliverables.\n- Inputs: Complete — The solver bundle provides valid train, validation, and test tables; six source groups; model and ensemble parameters; ten attack scenarios; policy settings; runtime metadata; and a non-answering triage utility.\n- Toolchain: Complete — Python 3.12 and NumPy 2.3.5 are free and obtainable, with compatible version pins and a single-machine CPU workflow requiring no network services.\n- Reference output: Complete — Visible and private gold artifacts cover every required schema, and the expert visible outputs match the visible gold files.\n- Rubric: Complete — Component weights, visible/private weights, numerical tolerance, pass threshold, partial-credit behavior, schema validation, and deterministic scoring logic are supplied and coherent.\n\n**File and Task Reachability:**\n\nAttachment set: `reachable`. End-to-end task: `reachable`.\n\nBoth `input_materials.zip` and `reference_output_and_evaluation_dependencies.zip` were inspected. Every archive member was inventoried and matched its embedded size and SHA-256 manifest. The review covered the method specification; CSV datasets; model, ensemble, attack, policy, task-card, validation, and rubric JSON; pinned runtime files; starter utility; execution wrapper; evaluator lifecycle and scoring source; provenance documentation; expert pipeline; all CSV and Markdown outputs; and every visible/private NPZ member. The archives were readable, unencrypted, bounded, and free of traversal paths, links, device entries, or unexpected executables.\n\nThe tables have complete schemas, unique IDs, finite feature values, binary labels, expected record coverage, and balanced source-program coverage. NPZ arrays use non-object numeric or Unicode dtypes and reconcile across attack identities, perturbation norms, defended predictions, reported metrics, intervention coverage, release gates, and decisions. No attachment member was omitted; uploaded programs were reviewed statically and not executed. Independent executable replay was not performed because review policy prohibits executing submitted programs and the review environment did not provide an approved Python runtime.\n\nThe completion path is clear: implement the documented CLI, run it on solver-visible inputs, produce the six audit artifacts alongside the reusable pipeline, then undergo exact visible scoring and an isolated unprivileged replay against independently provisioned inputs. Evaluator-only references remain separated from the solver workflow.\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. The evidence identifies the datasets as deterministic clean-room simulations but does not establish AI fabrication or a predetermined defect-removal fixture. No approved public-search capability was available; supplied provenance states that the task is newly authored and cites public research only as scientific context.\n\n**Recommended Changes:**\n\nThe submission provides a coherent professional workflow, complete evidence, deterministic references, and strong evaluation isolation. No changes are required.","reviewed_at":"2026-08-16T07:50:53.764963Z"},"claim":{"id":"4c0e7dfe-6288-4de0-a402-38303a76ee30","submission_id":"53adf6e2-ab55-4030-91ae-803073f1c7b9","submission_revision_id":"68b467da-9afe-4d8d-b6be-5c50c9722c81","builder":"builder-01","status":"active","claimed_at":"2026-10-15T15:04:26.780353Z","delivered_at":null,"ended_at":null,"delivery_note":null,"end_reason":null}},"revision_is_current":true}