v0.4.0 - renamed to EEGauge

EEGauge - new name, same project (formerly "BCI Dataset Cards").

Changed

• Renamed the project to EEGauge. The GitHub repo, the Python package, and the CLI
  command are now eegauge (the old GitHub URL redirects).
• Breaking: the command-line entry point changed from bcicards to eegauge, and
  the import package from bcicards to eegauge. Update any scripts or aliases.

Everything else is unchanged: honest dataset cards for public EEG/BCI datasets,
leakage-aware baselines, chance/significance, per-class metrics, and MOABB + EEGDash
metadata. 26 tests; CI green on Windows/Linux x Python 3.11/3.12. Not medical software.

v0.3.1 - EEGDash live-API fixes + real example

Patch release: validate the EEGDash backend against the live API and add a real example.

Fixed

• Map EEGDash dataset metadata to the real API field names: read the DOI from
  dataset_doi (stripping the doi: prefix) and the citation from author_year.
• Read BIDS entities (subject/session/task/run) whether nested under entities or
  flattened at the top level (the live find() records use the flat form).

Added

• A real example from a live scan of ds002718 (Wakeman face-processing, 18 subjects):
  examples/ds002718.md and examples/ds002718.provenance.json.

Verified

• 26 tests; ruff lint + format; CI green on Windows/Linux x Python 3.11/3.12.

v0.3.0 - EEGDash backend + evaluation provenance

v0.3.0 adds interoperability with EEGDash (the new OpenNeuro/NEMAR catalog) and an
evaluation-provenance record - complementary to EEGDash's discovery/loading, not a
duplicate of it.

Added

• A pluggable DatasetBackend protocol. The existing MOABB logic is exposed via a
  moabb backend; a new eegdash backend reads EEGDash metadata and record listings
  with no signal download.
• bcicards scan --backend eegdash <dataset-id> produces:
  • a metadata card (modalities, record/subject counts, license, DOI, BIDS validation
    status, and honest leakage risk factors - not a "RISK: HIGH" verdict), and
  • <id>.provenance.json (schema bci-evaluation-card/0.1) recording the exact
    cohort (subjects/sessions/records) so an evaluation can be reproduced.
• New optional card fields, an [eegdash] install extra, and a design spec at
  docs/design/eegdash-backend.md.

Scope / honesty

• The eegdash backend is metadata-only by design. No baseline is run for EEGDash
  datasets (they span many modalities/tasks); that is intentionally deferred.
• EEGDash is new and its API may change; the backend is isolated with a small surface.

Quality

• 24 tests (EEGDash client mocked - no network in CI); ruff lint + format enforced; CI
  green on Windows/Linux x Python 3.11/3.12. Still not medical software; no diagnostic claims.

v0.2.1 - review fixes (version reporting, docs)

Patch release fixing issues found in code review.

Fixed

• bcicards --version reported the wrong version. The version is now single-sourced
  from bcicards.__version__, used by both the CLI and pyproject.toml (dynamic
  metadata), so it cannot drift again. Covered by a new --version smoke test.
• Card plot paths use forward slashes, and the committed example card references its
  co-located images so they render on GitHub.
• Replaced non-ASCII em dashes in the README and .gitignore with plain ASCII.

v0.2.0 — smarter, more honest cards + more datasets

v0.2.0 makes the dataset cards smarter and more honest, and broadens dataset coverage.

Added

• Chance level + significance — cards report the chance level (1 / n_classes) and an
  approximate one-sided binomial test of whether the score beats chance (closes #1).
• Per-class metrics + confusion matrix — per-class precision/recall/F1 table and a
  confusion-matrix plot, so a multi-class result is interpretable beyond overall accuracy.
• License / DOI / citation surfaced in cards, read defensively from MOABB metadata (closes #2).
• More motor-imagery datasets — BNCI2014_004, Zhou2016, Weibo2014 (closes #3).
• .pre-commit-config.yaml (Ruff) and extended mocked tests for every new output.

Verified on BNCI2014_001 (subjects 1-3, leave-one-subject-out, seed 42)

• Accuracy 0.429 vs a 0.25 chance level (binomial p < 0.001) — but the per-class breakdown
  shows the baseline barely detects the "feet" class, detail a single accuracy number hides.

Honesty note

The "above chance" check is an approximate binomial test versus the naive chance level, not
a permutation test — a sanity check, not proof. Still not medical software; no diagnostic claims.

v0.1.0 — first public release

First public release of BCI Dataset Cards — a small, beginner-friendly CLI for
inspecting public EEG/BCI datasets and running honest, reproducible baselines.

Features

• bcicards scan — generate a Markdown dataset card from MOABB metadata.
• bcicards benchmark — run a CSP + LDA motor-imagery baseline (fixed seed), save
  JSON results and a metrics plot.
• Subject-aware (leave-one-subject-out) splitting for multiple subjects; stratified
  holdout for a single subject.
• Explicit leakage warnings and "what this does not prove" sections in every card.

Verified on BNCI2014_001 (seed 42)

• Within-subject (subject 1, stratified holdout): 0.751 accuracy
• Cross-subject (subjects 1-3, leave-one-subject-out): 0.429 accuracy
• Chance for 4 classes is 0.25 — the gap shows why the split method must be stated.

Quality

• Mocked unit tests; CI on Windows + Linux, Python 3.11 + 3.12.
• MIT licensed. Not medical software; no diagnostic or "mind reading" claims.