SensorFault-Bench 1.0.1

Patch release that makes dataset parsing and scoring independent of the
platform. All results in the paper and the paper artifact remain valid, since
the datasets and every reported value are unchanged.

Fixed

• Validation with --require-checksums accepts the downloaded ETTh1.csv on
  every platform. Every dataset CSV file is now parsed with correct rounding,
  whereas the default pandas parser rounds long decimals differently from
  platform to platform. This changes the recorded content checksum of ETTh1
  but not the values that models train on.
• scripts/preprocess_penmanshiel_hourly_wt08.py reproduces the published
  penmanshiel_hourly_wt08.parquet on every platform, so a rebuild from the
  raw SCADA files passes --require-checksums. Previously, the rebuilt values
  depended on the platform.
• Analysis and the fixed-channel-fraction evaluation read the logged per-sample
  errors exactly as written. Before, the last digit of full-precision outputs
  could differ between platforms, while reported values were not affected.
• A Git checkout on Windows keeps the committed bytes of paper_artifact/ and
  configs/baseline_hparams.yaml, so these files match the checksums that the
  manifests in paper_artifact/ record.

SensorFault-Bench 1.0.0

Camera-ready release for the NeurIPS 2026 Evaluations & Datasets Track paper
"Benchmarking Sensor-Fault Robustness in Forecasting". The results of the
default configuration are unchanged, as are its scored scenarios, selection
rules, and scoring.

Added

• paper_artifact/supplementary_tables/, the tables that the paper added in
  its camera-ready revision. Its README.md describes each table.
• data/croissant/, the Croissant metadata of the two curated datasets with
  their Responsible AI fields.
• The sensitivity scenarios negative_drift, negative_spike, and
  local_timing_warp.
• The severity sampling distribution beta_1_2, which
  PERTURBATION_SEVERITY_DISTRIBUTION selects. The default stays uniform.
• docs/benchmark_protocol.md, docs/sensitivity_analyses.md, a results
  schema and troubleshooting in docs/reproducing_benchmark.md, and paired
  transfer inputs in docs/extending_methods.md.
• CITATION.cff, this changelog, and continuous integration.
• docs/using_fault_operators.md, a guide to uses of the fault operators
  outside the benchmark, such as pretraining or stress tests calibrated to
  observed faults. The protocol section "Fault Exposure in Training" asks
  reports to name the fault operators that a model saw in training, including
  pretraining.
• notebooks/tutorial/quickstart.ipynb, a CPU tutorial that shows the scored
  scenarios, runs a small benchmark, and reads the reported quantities.
• configs/cli_help.yaml, the help text of every command-line option. The
  --help output of the three entry points groups the options by topic and
  shows their defaults.

Changed

• The README is restructured around the key findings, the results of the
  default configuration, and a guide to evaluating a new model or method.
  data/README.md adds download commands for the two curated datasets.
• The two curated datasets are published as Zenodo records with Croissant 1.1
  metadata, from which data/README.md downloads them. Their Parquet and CSV
  files are byte-identical to version 2 of the Kaggle records of the review
  period.
• The paper artifact is version 2,
  sensorfault-bench-neurips2026-paper-artifact-v2.
• The folders in paper_artifact/tables/ carry the name of their evaluated
  setting, for example default_configuration/ instead of
  primary_clean_2026-04-26_10k/. paper_artifact/MANIFEST.json records the
  export date and source path of each folder. The tables are unchanged.
• The figure scripts and PDFs in paper_artifact/R/ match the camera-ready
  figures. Their inputs are unchanged.
• paper_artifact/README.md maps each section of the paper to its bundled
  results and the commands that reproduce them, and states that the paper
  corresponds to release v1.0.0.
• The compute accounting moved from evidence/compute_accounting/ to
  paper_artifact/compute_accounting/. It also covers the seed reruns, ETTh1
  task settings, and sensitivity checks of the camera-ready revision. Its values
  for the default configuration are unchanged.
  scripts/compute_accounting_ledger.py writes it from the campaigns listed in
  configs/reporting/compute_accounting.yaml.
• The extension guides describe the interfaces, work through one example each,
  and list every configuration entry that a new architecture, method, or
  dataset needs. docs/extending_models.md shows how to evaluate a model whose
  weights were trained elsewhere.
• The architecture families of the analysis plots are configured in
  configs/benchmark_scope.yaml, and the trajectory legends use the method
  labels of configs/reporting/core_figures.yaml. Analysis therefore accepts
  new architectures and methods without code changes.
• instantiate_perturbations in data/data_module.py is now public.
• TSDataModule no longer accepts severity_laws.
  PERTURBATION_SEVERITY_DISTRIBUTION sets the severity sampling distribution.
• DatasetSpec requires input_channels.

Fixed

• Forecast replay binds the same randomized smoothing noise to each window as
  degradation scoring does. scripts/render_forecast_plots.py and the
  rendering of extreme forecasts in run_analysis.py therefore reproduce the
  scored forecasts of randomized smoothing runs instead of failing the score
  check.
• With STRICT_IID: true, every sampled window lies inside its own block
  together with its target horizon. No two windows therefore share a time step.
  The results of the default configuration are unchanged because the option is
  off by default and degradation evaluation rejects it.
• scripts/build_paper_artifact.py --force builds the new bundle in a staging
  directory and replaces the existing bundle only after the build succeeds. A
  failed build leaves the previous bundle unchanged.
• The artifact builder and the test suite run on Python 3.10, the minimum
  version declared in pyproject.toml. Continuous integration tests Python
  3.10, 3.11, and 3.12.
• Training, testing, analysis, and forecast replay stop when the input or
  target channel order recorded on a run differs from the dataset
  specification. The data configuration signature compares channel sets, so a
  run trained under another channel order was previously scored with the wrong
  channel mapping. Training also counted such a run as a completed candidate.
• Analysis and repeated testing accept runs that were reevaluated with fewer
  scenarios, as docs/sensitivity_analyses.md describes. They previously
  stopped on the metrics that the earlier evaluation had logged for the dropped
  scenarios.
• run_training.py trains baseline first and wrapping methods last, whatever
  the order of --method, so a method listed before baseline no longer fails
  on a new experiment. run_analysis.py stops before reading any runs and
  names the missing baseline when --method omits it.
• Training stops with a clear error when the validation split holds fewer
  windows than one batch. Validation scores full batches only, so such a split
  previously ended in an error about the missing early-stopping metric.
• Testing keeps the end time of each evaluated run. Because the end time breaks
  ties between equal validation scores, the new end time that evaluation used
  to record could make a repeated test select another winner after
  --refresh-selection-metrics.
• Scoring accepts bootstrap percentile intervals that exclude the point
  estimate, which a small BOOTSTRAP_CI_RESAMPLES can produce. Such intervals
  previously stopped scoring.
• Fault augmentation runs with data loader workers under the spawn start
  method, the default on macOS and Windows. Its severity sampler previously
  could not be passed to the workers.
• Randomized training stops when its noise level is not finite, for example
  nan. Such a run previously trained without noise.