Building time-series models, and looking inside them

LeNEPA is a compact representation model that drops the augmentation bag entirely. Aionoscope is the instrument we built to read how frozen time-series foundation models arrange signal geometry.

Series 1 · The model

LeNEPA

A small causal encoder that learns time-series representations by predicting its own next latent token — no augmentations, no bidirectional attention, no hand-built difference branch.

Representations without the augmentation tax

A 6.3M-parameter encoder lands within 0.24 points of a 161M-parameter baseline by dropping hand-tuned augmentations entirely.

Read article

Series 2 · The instrument

Aionoscope

Generate signals with known latent factors, freeze a foundation model, and read how it arranges phase, event time, and trend into geometric manifolds — across layers, scale, and training.

The Instrument

The generator, the frozen-model probe harness, and the manifold metrics — isometry, neighbourhood, projection, fiber — that score how faithfully a representation lays a factor out.

Read article

The Shape of a Signal

How phase, event time, and trend become geometric manifolds in the activations of a frozen time-series foundation model.

Read article