Domain shift is when new data stops matching what a model learned on, and it is where most models degrade. With an entire region held out of training, the Clone's 90% range still covered 90% of outcomes in that new region, the same as in-domain.
When an input falls outside what a Clone has seen, it abstains instead of guessing, and it says so. Coverage and the abstain rate are measured on held-out cases and travel with every prediction. Honesty is the point.
Accuracy is how close a prediction lands to the truth. On a public test a single model tracks the real answer 94% of the way (R2 0.94), within about 4% of the top published result, and it returns a calibrated range with every prediction.
Speed is how efficiently a model runs live. On the industry-standard harness the Clone served tens of thousands of predictions per second, on an ordinary computer with no GPU. Each model is kilobytes, small enough to run anywhere.
for the trust axis today's leaderboards do not measure:
whether a model stays honest when the conditions change.
