{"submission_id":"3e3ed5a5-ebd2-4e34-9519-8dd900c56a10","submission_revision_id":"a84d61ae-898b-4e17-a787-c750f3865582","revision_version":1,"submission_status":"under_review","submitted_at":"2026-08-29T16:29:42.366619Z","submitter_name":"Yunshi Zhang","submitter_email":"1785948079zys@gmail.com","task":{"status":"submitted","version":1,"extra_fields":{},"submitted_at":"2026-08-29T16:29:42.366619+00:00","industry_domain":"Computing, Data & Mathematical Sciences, Data & Analytics Engineering","task_short_name":"speech_audio_dsp_denoising_resampling","operating_system":"Linux","task_description":"Objective: Perform a deterministic speech and audio signal-processing experiment on the supplied synthetic speech-like waveform. Add controlled Gaussian white noise, design and apply the specified FIR denoising filter, perform decimation, zero-value insertion, and head/tail zero padding, then export reproducible audio files, numerical arrays, plots, metrics, source code, and an analytical report.\n\nStarting State: A Linux VM containing a deterministic 3-second mono PCM16 WAV file at 16 kHz, exact DSP and output contracts, a canonical plot renderer, starter code, a report template, and offline NumPy and Pillow dependencies.\n\nInstructions & Rules: Decode the supplied PCM16 signal exactly. Generate 6 dB AWGN using seed 20260829. Apply the specified 129-tap, 3500 Hz Hamming-window FIR low-pass filter with normalized DC gain and centered same-mode convolution. Perform direct decimation and zero insertion using factors 2 and 4, and add 4000 zeros to each boundary. Write the complete contracted output tree under `base/output/`. All arrays, WAV payloads, metrics, plots, report claims, and manifest hashes must agree with independent evaluator recomputation. Do not modify the input packet or claim microphone capture, human listening, perceptual testing, or speaker identification.","software_licensing":"In-house/Proprietary","evaluation_criteria":null,"verification_method":"Success is determined by twelve binary hard gates. Award a score of 1.0 only if every gate passes. If any submission-controlled gate fails, award 0.0. Missing or corrupted evaluator dependencies must be reported as an infrastructure error with a null score.\n\nInput immutability\n\nVerify that every file and directory under base/input/ remains byte-for-byte identical to the evaluator-owned input packet. Reject additions, deletions, modifications, symbolic links, and special files.\n\nExact output structure\n\nRequire exactly the seventeen contracted output files under base/output/: solution.py, metrics.json, analysis/results.npz, eight PCM WAV files, four PNG plots, REPORT.md, and result_manifest.json. Reject missing, unexpected, oversized, empty, unsafe, or escaping paths.\n\nManifest reconciliation\n\nVerify that result_manifest.json lists every other required output exactly once. Every relative path, byte count, and lowercase SHA-256 digest must match the final file.\n\nSource validation\n\nRequire solution.py to be valid UTF-8 Python source between 1,500 and 262,144 bytes. It must parse successfully and must be capable of regenerating the complete output from the supplied input.\n\nIndependent numerical recomputation\n\nSafely inspect analysis/results.npz before loading it. Require the exact ten float64 arrays, keys, shapes, and finite values. Independently recompute the original signal, seeded AWGN, noisy signal, 129-tap Hamming FIR filter, denoised signal, factor-2 and factor-4 decimation, factor-2 and factor-4 zero insertion, and 4,000-sample boundary padding. Compare all samples using rtol=1e-9 and atol=1e-10.\n\nMetrics reconciliation\n\nRequire the exact metrics.json schema. Recompute and compare all SNR, filter-response, sample-rate, sample-count, filter-parameter, and sampling-operation values. Reject duplicate keys, omitted fields, NaN, infinity, placeholders, and contradictory values.\n\nWAV validation\n\nDecode every submitted WAV and require mono, little-endian, uncompressed PCM16 encoding with the exact sample rate and sample count. Every PCM sample must equal:\n\nclip(round(array * 32768), -32768, 32767)\n\nPlot validation\n\nRequire four RGB PNG files of exactly 1400 by 900 pixels. Regenerate each plot from the validated arrays using the trusted canonical renderer and require exact decoded-pixel equality. Reject random, uniform, transparent, malformed, reused, or decompression-bomb images.\n\nReport validation\n\nRequire the six exact report headings in the specified order, substantive visible prose, a connected Mermaid processing flowchart, the published DFT/SNR/FIR/decimation/zero-insertion formulas, all four plot references, filter-design reasoning, and factor-specific aliasing and spectral-imaging analysis.\n\nReconcile all numerical and procedural claims with the validated arrays and metrics. Reject hidden placeholder content, contradictory filter or sampling claims, and unsupported claims about microphone capture, human listening, MOS, PESQ, STOI, subjective quality, or speaker identity.\n\nSignal-quality thresholds\n\nIndependently require:\n\nmeasured input SNR: 6 dB within 1e-9 dB;\n\nSNR improvement: at least 2.5 dB;\n\npassband ripple: at most 0.1 dB;\n\nstopband attenuation: at least 50 dB.\n\nClean public-input replay\n\nExecute the unchanged submitted solution.py in a fresh isolated directory using a pristine copy of the public input packet. Require a zero exit status, unchanged input files, bounded resource use, and a newly generated result tree that passes all applicable artifact checks.\n\nClean hidden-challenge replay\n\nExecute the same unchanged source again using a different evaluator-owned synthetic input with the same format and contract. Recompute all expected values for that signal and require another complete pass. Signal-dependent plots must differ from the public-input plots. This gate rejects solutions that merely unpack or hard-code the public reference output.\n\nOnly evaluator-generated computations and clean source replays count as evidence. Screenshots, copied logs, filenames, hashes without matching content, and self-authored pass statements cannot substitute for valid numerical artifacts.","software_and_version":"CPython 3.12, NumPy 2.5.2, Pillow 12.3.0, uv, using the version supplied by the ALE environment, ALE evaluation framework, Static Linux x86-64 Landlock/seccomp sandbox runner, 64-bit Ubuntu Linux on the ALE cpu-free-ubuntu snapshot","specific_requirements":null,"expected_completion_time":null},"files":[{"file_id":"ce8e6ece-4741-4079-a10d-b2b2a06b8044","filename":"speech_audio_dsp_denoising_resampling_Input_Materials.zip","category":"input_materials","content_type":"application/x-zip-compressed","size_bytes":23524204},{"file_id":"24332172-6027-403d-bb87-ca15dec48b97","filename":"speech_audio_dsp_denoising_resampling_Reference_Output_and_Evaluation_Dependencies.zip","category":"reference_output","content_type":"application/x-zip-compressed","size_bytes":107583231}],"deep_review":{"overall_verdict":"Strong Accept","review_text":"**Overall Verdict:** Strong Accept — The submission is complete, internally consistent, professionally scoped, objectively verifiable, and reachable end to end without material contributor-side gaps.\n\n**Completeness Check:**\n\n- Instruction: Complete — Exact DSP procedures, numerical tolerances, output structure, reporting requirements, and prohibited claims are unambiguous.\n- Inputs: Complete — The deterministic PCM16 WAV, provenance, contracts, renderer, starter materials, dependency wheels, and integrity manifests are present and consistent.\n- Toolchain: Complete — A hash-declared CPython 3.12 runtime, compatible offline NumPy and Pillow wheels, Linux sandbox runner, adapter, and licensing notices provide a viable toolchain.\n- Reference output: Complete — The 17-file known-good output is substantive; arrays, WAV payloads, metrics, plots, report, and manifest reconcile with direct non-executing checks.\n- Rubric: Complete — Twelve explicit binary gates cover input integrity, artifacts, numerical correctness, reporting, quality thresholds, and public and hidden replay.\n\n**File and Task Reachability:**\n\nAttachment classification: `reachable`. Task classification: `reachable`. Both `speech_audio_dsp_denoising_resampling_Input_Materials.zip` and `speech_audio_dsp_denoising_resampling_Reference_Output_and_Evaluation_Dependencies.zip` were fully inspected. Their 22-member and 131-member inventories are readable, unencrypted, non-traversing, and integrity-consistent. Review covered contracts, manifests, UTF-8 and JSON structure, WAV encoding, NPZ schemas and transformations, PNG dimensions and representative visual content, sanitized source pages, reports, evaluator source, dependency metadata, and nested runtime topology. No task-relevant attachment content remained unreviewed. Uploaded programs and binaries were inspected statically but not executed; clean replay was not independently run because evidence-review policy prohibits executing uploaded code and no independent compatible Python runtime was available. The completion path is clear: process the public WAV using the supplied contracts, generate the exact output tree, validate it through independent recomputation, then replay unchanged source against pristine public and hidden inputs under the bundled isolation path.\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\n\n**Recommended Changes:**\n\nThe submission already provides unusually thorough contracts, integrity controls, reference evidence, hidden-challenge protection, and negative-test coverage. No changes are required.","reviewed_at":"2026-08-29T16:40:56.001444Z"},"claim":null}