Active Learning

Loop bookkeeping (phase 1: campaigns, immutable checksummed candidate pools, human-closed cycles), deterministic acquisition selection (phase 2), the human review step (phase 3: review decisions, human-supplied experimental observations, immutable labeled dataset-version commit), and explicit human-triggered retraining (phase 4: derived readiness, one explicit Retrain command through the canonical B21 training engine, immutable result). A selection is a computational choice derived from B21 model predictions — always a DERIVED_RESULT, never an experimental measurement and never a claim of scientific optimality. A label is a human-supplied experimental observation; a model prediction is never converted into a label. Retraining and evaluation are COMPUTED results — nothing here claims a retrained model is better than its parent, that Active Learning improves performance, or that any candidate is experimentally validated. Nothing retrains automatically: retraining fires only on an explicit human action.