seismic-snr-filtering-bias

2026-09-30 archive correction

The dated correction adds 72 existing integer TP/FP/FN arrays that were incorrectly excluded as waveform caches. Their archived totals reproduce the saved scores; they are not complete probability traces. No scientific results changed. The current gap list remains nonempty and specifies prediction, historical model-identity and clean-start limits. Raw datasets are cited, not redistributed.

Waveform contrast does not measure training value in seismic deep learning

Reproducibility archive for the manuscript by Ziye Yu, Xin Liu and Yuqi Cai. This is a research artifact repository, not a claim that the manuscript has been published or accepted.

The 2026-09-29 NC revision adds the available project-controlled materials for the extended comparisons in Supplementary Notes 12–18. Original fixed-endpoint experiments remain at the root. Use an immutable repository commit when citing or downloading; the paper records the verified commit. Repository DOI: 10.57967/hf/9115. The manuscript's 10.57967/hf/9778 also resolves to this repository; neither identifier is a new DOI for this deposit.

Current entry points

• Coverage and remaining limitations
• Data sources and split definitions
• Reproduction guide
• Artifact index, file-level bundle manifests, and SHA256 checksums
• Frozen source and execution records
• Audit ledgers, score tables and figure derivatives

The added bundles contain split/selection manifests, trained checkpoints and saved states, validation histories, derived test-window counts or path-level error arrays, monitoring candidate predictions, selected streams, associations, matching outputs and dependence analyses. Each tar member has a size and SHA256 entry. Full probability traces and full period-by-period predictions are not available for every learning comparison; aggregate scores or sufficient statistics must not be described as those complete predictions.

Raw upstream datasets are not added. Obtain CREDIT-X1local, SeisDispFusion-NCF and SeismicX-Cont from their cited providers under their terms. Waveform-containing evaluation caches and original restricted split-key files are excluded. Derived record IDs, manual-label metadata, SNR values and membership manifests are supplied to document the splits. Monitoring JSONL files are model predictions with source/station metadata, not waveform samples.

Verify and unpack

After downloading this repository at a fixed commit:

python nc_revision_20260929/verify_release.py --repo .
python nc_revision_20260929/prepare_workspace.py --repo . --out ../snr-reproduction

The first command checks SHA256 for every listed file. Add --members to also decompress and validate every archived member. The second creates a fresh workspace and paper-driver directory without changing frozen source bytes. It refuses an existing output directory and never downloads raw data. See the guide before training: some frozen drivers retain machine-specific paths, and a fresh end-to-end run has not been certified for this deposit. The pre-deposit missing-entry checks are retained as historical observations and are not reclassified as successful runs by the act of uploading files.

Original archive

The original material remains in code/, checkpoints/, training_manifests/, configs/, and results/. The prior revision was da82ce1fe8b571a43c93a91df55d60d6ddcac8e3. Its README, open-data guide, and crosswalk describe that historical layout. Historical manuscript/figure labels are not current SI numbering.

Project-controlled materials use CC BY 4.0. Third-party data and components retain their own terms; the included REAL source has its MIT license beside it. Cite upstream datasets as well as the repository revision.

Reviewer verification added 2026-10-01

Executable verification package and core-result crosswalk provide seven actual result-reconstruction checks, including the minimal ordered reference ID/hour input for exact bootstrap reproduction. The frozen result bundles are unchanged. These checks do not certify a complete fresh waveform-to-training pipeline; historical picker identity and full probability/period-output limits remain disclosed.