Inside the MetaModel: ensembles over intuition_

The MetaModel is not one model. It is an ensemble of deliberately simple learners — gradient-boosted trees over engineered features, per-participant calibration curves, and sequence models over session history — combined by a stacking layer that is retrained nightly.

Why not one big model

With a low signal-to-noise ratio and a dataset measured in tens of thousands rather than billions, large end-to-end architectures overfit long before they generalise. Weak learners with strong priors are the honest choice at this scale.

Ensembles also fail gracefully. When one component degrades — a target pool changes, a cohort churns — the stack absorbs it instead of collapsing.

Closing the loop

Model output returns to participants as personal calibration feedback. Better-calibrated participants produce cleaner data, which produces a better model. The loop is the product.