The verifiable-training-mechanisms programme asks how to pay untrusted workers to train a model when the only check is re-running one step within a tolerance. This session drafted the paper body from the programme's log.
The measured result: a worker that skips every step the referee cannot tell from honest work passes 30 to 67% of steps at a tolerance of 1e-2, about the worst float16 drift. The held-out loss it gives up there is under one seed standard deviation on digits and negative on breast cancer (skipping late, small steps acts like early stopping; it won on 19 of 20 resampled splits). Loss is lost only at 2 to 6 times the worst benign drift.
Two conjectures failed and are in the paper as negative results. The stake that deters a majority coalition is max(0, g/p - 1), flat in the number of workers (140 cells, matching an exhaustive search to 7e-10), not falling as 1/m. A first-order predictor of the degradation misses by 3 to 100 times.
Limits: three small scikit-learn tasks, plain SGD, no transformer yet, and the draft is uncompiled. Source: research/verifiable-training-mechanisms/.