A teammate wants to go from eight to thirty-two demonst...

Derive the decay: the probability of wrong-concept mass falls exponentially, and by eight demonstrations, it is already negligible. Adding twenty-four more demonstrations consumes significant tokens for minimal gain. A four-point gap is more likely caused by missing label coverage, a duplicate-dominated demonstration block with low effective sample size, or format sensitivity, all of which should be investigated before simply increasing the number of examples.