TrustLens v0.4.0 — Framework-Agnostic Trustworthiness Platform

🚀 TrustLens v0.4.0 — Framework-Agnostic Trustworthiness Platform

TrustLens v0.4.0 is a major architectural milestone that transforms TrustLens from a scikit-learn-focused extension into a framework-agnostic platform for ML trustworthiness evaluation.

This release introduces a new backend resolver architecture, native XGBoost support, improved auditability, stronger calibration correctness, and substantial production hardening — all while maintaining zero breaking changes for existing scikit-learn users.

---

🏛️ Framework-Agnostic Prediction Resolver Architecture

At the heart of v0.4.0 is a complete decoupling of prediction generation from trustworthiness evaluation.

TrustLens now uses a Prediction Resolver Architecture that:

• Automatically detects supported ML frameworks
• Resolves predictions into a standardized internal contract
• Preserves framework-specific metadata for auditing
• Enables future backend expansion without modifying core analysis logic

This lays the foundation for broader ecosystem support while keeping the analysis engine framework-agnostic.

---

🌟 Highlights

✅ Native XGBoost Support

TrustLens now supports:

• XGBClassifier
• raw xgboost.Booster objects

Features include:

• automatic DMatrix handling
• probability normalization
• multiclass support
• objective-aware task blocking (reg:*, rank:*)

Install optionally:

pip install "trustlens[xgboost]"

---

✅ Manual Prediction Pipelines (model=None)

TrustLens can now audit external inference pipelines and unsupported frameworks using:

analyze(
    model=None,
    y_pred=predictions,
    y_prob=probabilities,
)

This enables:

• servability audits
• external model validation
• offline prediction evaluation
• framework-independent trust analysis

---

✅ Scientifically Correct Multiclass Calibration

Fixed a significant calibration issue affecting multiclass classification.

TrustLens now correctly computes the Multiclass Brier Score using class-wise mean squared error, instead of incorrectly assuming binary probabilities.

This improves the scientific reliability of calibration metrics for complex classification systems.

---

✅ Audit Provenance & Metadata

Every TrustReport now includes backend provenance:

• framework detection
• resolver metadata
• framework versions
• manual override tracking
• degraded analysis transparency

This improves reproducibility, experiment tracking, and deployment auditing.

---

✅ Stability & Hardening

v0.4.0 includes substantial production hardening:

• NaN/Inf validation
• probability range enforcement
• EPS-based numerical tolerance
• automatic probability clipping
• robust override handling
• graceful degraded-mode execution

When probabilistic outputs are unavailable, TrustLens now explicitly records:

{
  "degraded_mode": true,
  "missing_components": [...]
}

for full transparency.

---

📦 Unified Export Artifacts

Trust reports are now easier to persist and consume.

report.save("report.json")

generates a single, self-contained artifact containing:

• results
• metadata
• trust score
• backend provenance

Perfect for:

• CI/CD model gating
• experiment tracking
• ML observability
• deployment audits

---

🛡️ Compatibility & Migration

Zero Breaking Changes

Existing scikit-learn workflows continue to work unchanged.

No migration is required for current users.

Optional XGBoost Dependency

XGBoost remains fully optional.

If unused, TrustLens will not import or require it.

---

📈 Release Quality

v0.4.0 ships with:

• 263 passing tests
• characterization parity testing
• backend stress testing
• packaging validation
• production-like smoke testing
• ~75% test coverage

---

📚 Documentation

• Full Changelog → CHANGELOG.md
• Architecture → docs/architecture.md
• Backend Contract → docs/internal/prediction_contract.md

Thank you to everyone following and contributing to TrustLens 🚀

This release establishes the foundation for future backend support across the broader ML ecosystem.