Post-sampling control of continuous attributes

Conditional generators do not always produce samples that match the requested value of a continuous attribute. In 12-lead ECG synthesis, raising the requested age from 20 to 80 years raises the predicted age of the generated ECGs by only about 35 years, so synthetic cohorts carry age labels that their waveforms do not express. Rather than retraining the generator, we correct its samples afterwards: an assessor estimates the attribute a sampled latent expresses and a bounded residual editor trained on real latents moves the latent towards the request, with losses for locality, reversibility, label preservation and realism. Applied to age in PTB-XL ECG synthesis, this calibration reduces age error under an external ECG-age model by about 0.9 years per sample and by 2 years over 20 to 80-year requests, outperforms input remapping and guidance changes and leaves diagnosis AUROC nearly unchanged.