What accuracy can a model reach when trained on a synth...

A strong answer avoids the simplistic ‘capped at teacher accuracy’ response and instead decomposes the error. By using repeated sampling, you can separate errors that recur from those that vary. The student converges to the mode, setting a bound of 1 - ϵsys. Because this bound is abstract, you must perform a small gold audit over multi-draw outputs to measure the actual noise mass. A weak answer fails to account for the conditions under which students can exceed teacher performance.