GuardRail Neural InspectorKeras Tensor Engine
In-Browser Dense(96) → Dense(32) → Sigmoid Inference · Trained on UCI SMS Spam Corpus
Loading weights...
Corpus Benchmark
UCI SMS Spam
Held-Out Accuracy
...
Macro F1 Score
...
Exported Vocab
...
Real-Time Sigmoid Risk Telemetry
Unsafe / Spam Posterior P(y=1|x)0%
0% (Clean)Review Gate (25%)Block Gate (65%)100%
Dense Layer 1 (ReLU)
0 / 96 active
Dense Layer 2 (ReLU)
0 / 32 active
Sigmoid Logit Gate
0.0000
Evaluation Split Metrics
Loading Keras model artifacts...
Token-Level Risk Attribution
Each matched vocabulary token is probed individually through the exported Keras weights to rank its contribution toward the spam/unsafe class:
No known vocabulary tokens found in current input.
Recent Audit Snapshots
1. Regex tokenizer extracts lowercase alphanumeric tokens
2. Log-normalized BoW vector mapped over 1,600-token vocab
3. Forward pass: Dense(96, ReLU) → Dense(32, ReLU) → Sigmoid(1)
4. Calibrated threshold gate routes to Allow / Review / Block