Explainability
What the MITCF models can and cannot explain about their predictions.
Techniques Not Applicable to This Implementation
✗
SHAP / LIME: Requires NumPy/scikit-learn compatibility — not available in the pure-Python implementation
✗
Grad-CAM / Attention Maps: No deep learning framework — forward pass is manual Python loops with no autograd
✗
Saliency Maps: Requires differentiable computation graph — not present in pure-Python MLP
✗
Feature Importance (permutation): Possible in principle but computationally expensive with 64 hashed features and no framework
✗
Layer Activation Analysis: Manual forward pass does not store intermediate activations in an accessible form
Available Explanations
Token Presence
SHA-1 hash of individual tokens mapped to feature bins. Most informative features are high-frequency terms in the benchmark corpus.
Token Frequency
Counts of hashed token occurrences. Captures repetition patterns correlated with hallucination risk.
Bigram Hashes
Consecutive word pairs. Captures phrasing patterns associated with trustworthy vs untrustworthy outputs.
Fuzzy Gate Interpretability
The FuzzyDecisionGate is fully interpretable by design. Each decision is explained by:
✓
5 Crisp Input Values
quality, safety, risk, uncertainty, efficiency — each computed from model metrics and mapped to [0,1]
✓
Membership Degrees
Each input's degree of membership in LOW/MEDIUM/HIGH fuzzy sets, visible per algorithm
✓
15 Rule Activations
Each Mamdani IF-THEN rule contributes a weighted vote for its output action (see Fuzzy Rules page)
✓
Centroid Gate Score
Final crisp score via centroid defuzzification; decision = nearest action centroid