Skip to content

CHANGELOG.md

Latest commit

 

History

History
72 lines (60 loc) · 4.31 KB

CHANGELOG.md

File metadata and controls

72 lines (60 loc) · 4.31 KB
Raw

Changelog

[0.3.0] - 2026-04-01

Added

  • New audit/ subpackage: explainability audit trail for insurance pricing models.
  • ExplainabilityAuditEntry: tamper-evident dataclass capturing one model prediction event. SHA-256 hash of all fields except entry_hash enables immutability verification. Fields include SHAP feature importances, raw prediction, final premium, human reviewer identity (SM&CR CF reference), override flag and reason, and decision basis.
  • ExplainabilityAuditLog: append-only JSONL log. Supports read_since() filtering, verify_chain() hash integrity check across all entries, and export_period() for regulatory submission with a metadata header line.
  • SHAPExplainer: wraps the optional shap library. Supports tree, linear, kernel, and deep explainer types. Returns signed SHAP values as plain dicts keyed by feature name. shap is an optional dependency; raises ImportError with clear install instructions if not present.
  • PlainLanguageExplainer: converts SHAP values to plain English sentences suitable for FCA PRIN 2A Consumer Duty customer communications. Supports GBP/EUR/USD, per- factor pound impact scaling, override and rule-fallback notes, and bullet-list output.
  • AuditSummaryReport: builds HTML and JSON audit summaries covering decision volume, feature importance distribution (mean absolute SHAP), human override rates, per- segment analysis, and hash integrity status. HTML is fully self-contained.
  • Top-level re-exports for all five audit classes added to insurance_governance.
  • 65+ tests in tests/test_audit.py covering entry creation and hash verification, log operations, plain language output, report generation, and SHAPExplainer with mocked shap library.

Changed

  • Version bumped from 0.2.0 to 0.3.0.
  • pyproject.toml keywords updated to include explainability, SHAP, audit trail.

[0.1.10] - 2026-03-31

Changed

  • Corrected SS1/23 regulatory framing throughout: SS1/23 is a banking supervisory statement and does not apply directly to Solvency II insurers. All documentation now uses "aligned with SS1/23 best practice" language rather than implying direct applicability.
  • Added Consumer Duty (PRIN 2A) and TR24/2 as the primary mandatory regulatory hooks for GI pricing model governance in README and governance pack template.
  • Added PRA SoP3/24 (IMOR annual attestation) as the PRA-side governance expectation in README regulatory framework table.
  • Updated pyproject.toml description and keywords to reflect Consumer Duty, TR24/2, SoP3/24, and IMOR rather than SS1/23 as the primary framing.
  • Updated report.html.j2 template: regulatory reference table now cites Consumer Duty (PRIN 2A) + TR24/2 for performance validation and model inventory; SS1/23 references replaced with "good MRM practice (aligned with SS1/23 best practice)" framing; report header and footer updated.
  • Updated CONTRIBUTING.md: clarified that SS1/23 citations in code should use "aligned with SS1/23 best practice" language.
  • Updated source module docstrings to use consistent "aligned with SS1/23 best practice" phrasing.

[0.1.5] - 2026-03-23

Fixed

  • Bumped numpy minimum version from >=1.24 to >=1.25 to ensure compatibility with scipy's use of numpy.exceptions (added in numpy 1.25)

v0.1.3 (2026-03-22) [unreleased]

  • fix: correct license badge (BSD-3 -> MIT) and add missing Homepage URL
  • fix: use plain string license field for universal setuptools compatibility
  • fix: use importlib.metadata for version (prevents drift from pyproject.toml)
  • fix: correct Model C PSI verdict and benchmark filename reference

v0.1.3 (2026-03-21)

  • Add cross-links to related libraries in README
  • docs: replace pip install with uv add in README
  • Remove duplicate CTA from README
  • Add blog post link and community CTA to README
  • Add MIT license
  • Add discussions link and star CTA
  • Add real benchmark numbers to Performance section (Databricks 2026-03-16)
  • Remove pydantic v2 dependency for Databricks serverless compatibility
  • Fix regulatory accuracy: SS1/23 applies by analogy to insurers, not directly
  • Add PyPI classifiers for financial/insurance audience
  • Add Colab quickstart notebook and Open in Colab badge
  • Fix P1 bugs: exposure in A/E ratio, HL degrees of freedom, tier assignment order, Tier 4 dead code
  • Fix docs workflow: use pdoc not pdoc3 syntax (no --html flag)