Scaling Laws in Reverse: How Proxy Cascades Predict Billion-Parameter Outcomes

The most expensive way to learn whether a dataset can support a large model is to train the large model. Scaling-law research gave the field a better idea: many properties of training behave predictably across model sizes. If a signal degrades smoothly as you shrink a model, you can often run the movie in reverse — measure small, predict large.

Synthos operationalizes that idea for data quality. Each validation trains a cascade of 15 to 18 proxy models spanning roughly 1M to 500M parameters on stratified samples of the candidate corpus, then fits how each quality dimension trends across scale.

A single proxy model tells you how one point in scale-space responds to your data. It cannot tell you the direction of travel. Collapse-prone data often looks acceptable at 10M parameters and only reveals its ceiling in the curvature between scales — the gap between where the fit says quality should land and where it actually lands.

The cascade also lets us isolate which dimension is dragging the extrapolation. A corpus can have excellent distribution fidelity and still fail on temporal consistency; the fix recommendations differ completely, and a one-model probe cannot separate them.

When we say Synthos predicts billion-parameter outcomes with 90%+ accuracy, we mean the extrapolated risk assessment agrees with observed full-scale behavior within stated confidence intervals — and we publish those intervals on every report rather than a bare point estimate. The warranty program depends on this honesty: we financially back predictions, so we cannot afford to flatter them.

The cascade design also delivers a 49% efficiency gain over the traditional alternative of dress-rehearsal training runs, which is what makes running validation on every corpus — not just the scary ones — economically sane.