Overview

TrustLens is a reliability-focused evaluation layer for classification models. It helps teams move from metric reporting to deployment decisions backed by evidence.

Why This Matters

Accuracy alone is often insufficient for production decisions. A model can score high on accuracy while still being unsafe to deploy because of:

  • overconfident errors

  • subgroup performance disparity

  • weak probability calibration

TrustLens addresses this by combining diagnostic modules and explicit decision logic.

What TrustLens Evaluates

TrustLens evaluates models across four dimensions:

  • Calibration: are predicted probabilities aligned with real outcomes?

  • Failure behavior: are errors concentrated in high-confidence regions?

  • Bias and fairness: do important subgroups see uneven performance?

  • Representation quality: are embeddings well separated when provided?

These diagnostics are combined into a Trust Score, with penalties and blocker rules applied for high-risk conditions.

Typical Workflow

  1. Run analyze(model, X_val, y_val, y_prob=...).

  2. Inspect the returned TrustReport.

  3. Review score, blockers, and dimension-level outputs.

  4. Export artifacts for CI, governance, or comparison.

What You Get

A TrustLens run produces:

  • module-level metrics

  • a composite Trust Score with grade and verdict

  • narrative insights and detected risk patterns

  • saveable artifacts for downstream workflows