Atelier v0.5.1 — maxsim cosine and a multiclass SVM head.

Atelier v0.5.1 ships on Cloudera AI with a tighter architecture of record. The DST cosine channel is renamed to maxsim — a roll-forward without alias — to reflect the late-interaction implementation that scores a column’s values against vocabulary anchors token-by-token rather than against a single pooled embedding.

The SVM head moves to a Crammer–Singer multiclass formulation that learns one joint margin instead of one-vs-rest, and the UAT line is consolidated to a single calibrated path so the evidence channels coming out of training are the evidence channels going into the classifier.

Subsequent commits ground the test fixtures in CCO ExtendedRelation annotations and frame coverage as a correctness invariant — the Mars Climate Orbiter failure mode applied to ontology grounding.

Read the Atelier docs or the project page.