Signals: DST evidence fusion expands to five sources.
The Signals classification pipeline now combines five evidence sources. The new short-text SVM head reads character n-grams — sparse lexical features with no dependency on the sentence-transformer embedding — and directly addresses the source-independence concern in Dempster’s rule.
The release also lands a bootstrap agent that performs table-aware
batching, data-element discovery, and schema introspection before
classification begins, so the downstream pipeline isn’t starting
cold. SHAP explanations expose which slice of the 992-dimensional
feature vector mattered for each decision, and the synthetic-training
path now ships with a --self-train mode that reproduces LLM
annotations at 99.4% accuracy.
The 74-scenario BDD framework passes in air-gap mode with zero external network calls. See the Signals project page.