Changelog 0.8

Release notes for nimbus-bci 0.8.x and 0.7.x. The full historical log is shipped in the PyPI source distribution (CHANGELOG.md). Upgrading from 0.6 or 0.7? See Migration to 0.8.

---

​

[0.8.2] — 2026-10-16

​

Added

• Live Safety Dashboard (examples/trust_dashboard.py): One-command streaming dashboard visualizing verified rolling accuracy, regret CUSUM, mean/cov drift statistics, accept rates, governor commits, and budget exhaustion. Supports terminal and graphical render ([viz] extra).
• Wallet Profile Demo (examples/wallet_profile_roundtrip.py): Tap-to-Roam demo packing fitted user profiles into <30 KB pass-shaped bundles (neural_profile.pass/) with SHA-256 manifests.
• Lab Streaming Layer Ingest (examples/lsl_live_loop.py + [lsl] extra): Real-time EEG and marker inlet ingest with timestamped ring buffer and event-locked epoching. Includes --simulate flag to test without physical hardware.
• Preregistered Leaderboard Generator: Generated headline leaderboard script (scripts/generate_leaderboard.py) pulling directly from diff-protected golden fixtures.

​

Fixed

• README and quickstart direct construction examples aligned with DecisionConfig(preset=...) signature.

---

​

[0.8.1] — 2026-10-16

​

Fixed

• adapt_pseudo mid-batch cap refusal accounting: Commits accepted trials to the label budget and writes pseudo_report_ before raising AdaptationRefusedError.
• Ordered format checking in Personalizer.load: Older format versions load with a UserWarning rather than hard-failing with misleading “written by newer SDK” messages.
• restore_head trunk-generation warning: Emits a warning when restoring a head snapshot that predates the current live trunk geometry.

---

​

[0.8.0] — 2026-10-14

​

Added

• Public Safety Module (nimbus_bci.safety):
  • LabelBudget: Session and lifetime label spend ledger with monitoring floor exemptions.
  • AdaptationGovernor: Burn-in, cooldown, and adaptation count state machine with pure check() and atomic commit().
  • DriftMonitor: Persistent run-length mean and covariance shift surveillance.
  • SafetyMonitor: Single-arm regret CUSUM watchdog with Monte-Carlo calibrated thresholds.
  • AdaptationRefusedError: Raised with refusing control name, current state, and actionable remedy.
• Opt-in Safety Enforcement: Personalizer(..., safety=SafetyConfig(...)) gates all online adaptation sites (partial_fit, adapt_pseudo, adapt_trunk, adapt_async, adapt).
• Head Snapshot & Exact Rollback: Personalizer.snapshot_head() -> HeadSnapshot and Personalizer.restore_head(snapshot).
• Epochs Ingest Contract: Detailed ingest documentation for event-locking, from_mne_epochs, sampling rate discipline, and baseline correction.

​

Changed

• adapt_pseudo hardened defaults: Entropy ceiling ($0.60 \cdot \log_2 K$) and epistemic uncertainty ($0.45$) enabled by default; max_pseudo_updates=200 cumulative cap; fails closed on heads without MC uncertainty.

---

​

[0.7.3] — 2026-10-14

​

Added

• Balanced Accuracy Metric: Added nimbus_bci.metrics.balanced_accuracy and populated EvaluationResult.balanced_accuracy / CrossValResult.balanced_accuracy by default for rare-target paradigms (P300 / RSVP).
• Deployment Examples: Added deployment_fit_calibrated.py and embeddings_contract.py.

---

​

[0.7.2] — 2026-10-14

​

Fixed

• Honest Joint-Gate Calibration: calibrate_decision_gate and fit_calibrated now calibrate against the joint gate (confidence + entropy ceiling), eliminating undershoot at $K > 2$.
• De-biased Geometry Recommendation: recommend_geometry uses repeated stratified shuffle-split CV (n_repeats=24) to eliminate winner’s curse holdout inflation.
• Active Learning Hardening: CalibrationSession copies pools at construction, checks for pool contamination, and ensures idempotent stopping histories.

---

​

[0.7.1] — 2026-10-14

​

Fixed

• Digest-Bound Session Bundles: MISessionBundle and P300SessionBundle bind artifacts with format-2 SHA-256 digests (profile_sha256, sidecar_sha256) to prevent tampering and cross-bundle contamination.
• fit_population and adapt_trunk track embedding width across adaptation cycles and resets.

---

​

[0.7.0] — 2026-10-12

​

Added

• Precomputed Embeddings Input Contract (x_kind="embeddings"): Allows Personalizer to consume embeddings directly without calling encoder.encode.
• PersonalizerInputError: Clear exception raised when attempting trunk-level operations on precomputed embeddings.
• Full contract round-trip persisted in profile metadata.

---

​

Next Read