Watch the call flow

Demo scope: the RoBERTa intake stage is omitted from this visualization because the post-B2 trained model collapses onto near-uniform output for short prompts (val_loss 5.6088, underfit). Synthetic intake records drive the predictor + router below, which is honest about where the v3 work delivered measurable gains. See the router model card for the +1.27 reward learning curve.