Skip to content

Reclassify a pattern

The Studio’s classify strip is a decision, not a prediction. Atlas’s heuristic classifier picks a pattern; you can disagree and change it. The Reclassify button is right on the strip.

Anatomy of the classify strip CLASSIFY Analytical consumption view 0.84 keyword coverage · heuristic I_Customer · I_SalesOrder Reclassify ↑ what Atlas chose ↑ score ↑ honesty chip: this classifier is heuristic, not probabilistic ↑ the references Atlas leaned on ↑ disagree? the rest of the pipeline re-runs A decision, not a prediction. Something has to be committed or the planner has no ground to stand on.
The classify strip — five elements on a single line, each one reviewable.
  • The score is low or the keyword-coverage badge shows thin evidence.
  • The referenced artifacts the classifier cited are not what you meant.
  • You wrote a transactional task (RAP behavior, workflow trigger) and the strip picked analytical, or vice versa.

Classifying correctly before planning costs a few seconds; planning against a wrong shape and re-doing everything downstream costs an hour.

It is the single-line band across the top of the Studio canvas. When a task has been run, it shows: pattern name · score chip · keyword-coverage badge · reference artifacts · Reclassify button.

A menu opens with the available patterns. Pick the one that matches your intent. Atlas calls POST /api/atlas/studio/classify with the override and re-renders the strip.

Reclassifying invalidates the rest of the pipeline. Atlas:

  • throws away the existing task graph (its ancestry is kept on the evidence node so you can diff later)
  • re-runs the planner against the new pattern
  • regenerates any scaffolds the plan requires
  • re-runs the validation gates

The graph canvas re-draws once the new plan lands. Evidence rows update in the inspector.

Atlas stores the reclassification as an event on the case’s evidence timeline — the actor, the timestamp, the old decision, the new one. When the override matters for review, add a comment on the case that explains why; the comment stays attached to the same evidence row.