DQ Sentry v0.1.0

DQ Sentry v0.1.0

First public release of an explainable, local-first training-data quality and
debugging system for dataset maintainers.

Highlights

• Modular detectors for label issues, exact and semantic duplicates,
  train-test leakage, anomaly/poison risk, influence, data valuation, and
  weak-supervision disagreement.
• Budgeted, diversity-aware review selection and a local review dashboard.
• Reproducible evaluation, versioned manifests, runtime/memory instrumentation,
  Spark exact-duplicate profiling, CLI workflows, tests, and Docker targets.
• Public evidence for BANKING77 and aggregate-only evidence for AG News.

Evidence boundaries

The AG News blinded review involved one independent non-domain human reviewer:
84 confirmed, 8 dismissed, and 8 needs-context; 91.3% adjudicated precision
(Wilson 95% CI 83.8%–95.5%), with 50 minutes total and a 30-second median.
This is not expert adjudication, authoritative correction, inter-rater
agreement, whole-queue precision, external adoption, or proof of time saved.
No broad downstream-improvement claim is made.

The upstream BANKING77 pull request remains open and unmerged.

Privacy and licensing

This release contains no downloaded datasets, raw AG News text, blinded review
packets, private unblinding keys, raw reviewer responses, or reviewer-identifying
artifacts. Reproduction instructions download source data directly.

Release artifacts contain the Python wheel, source distribution, and checksums
only. No dataset artifacts or hosted service are published.