Logo
MAF
MAFStatistics

ML Analysis

A lightweight ML model that learns online from the LLM tier's verdicts and distils them into a microsecond classifier. Its decisions run in shadow mode.

Bootstrap (offline distillation)
Shadow mode: every ML decision below is recorded and scored against the LLM, but it is not used to block or allow any request. This lets you watch the model evolve before it is trusted to decide.
Agreement with LLM
-
how faithfully the ML distils the LLM
Accuracy vs ground truth
-
ML verdict vs the true attack/benign label
Decisions learned
-
LLM verdicts the model trained on
Model updates
-
online SGD steps applied

Learning curve

agreement with the LLM as it sees more decisions

Confusion vs LLM (teacher)

True Block
-
LLM & ML both BLOCK
Missed Block
-
LLM BLOCK, ML ALLOW
Extra Block
-
LLM ALLOW, ML BLOCK
True Allow
-
LLM & ML both ALLOW

Recent shadow decisions

RequestLLMML (shadow)ScoreMatch