Blog

Research &
Notes

Research2026-07-05

Reading the interval: transcription from the space between characters

Entreacte learns interval-glyphs — the space between two characters, indexed by the pair that bounds it — rather than the characters themselves. On regular-form text, a classifier identifies each junction and the overlap between intervals reconstructs the text at ~1.7% character error, OCR-grade, using structure alone with no language model.

Research2026-04-10

Why reversing binarization changes everything

Conventional OCR treats background as noise to be eliminated. We treat it as signal. Here's the mathematics behind the approach and why it lets the pipeline work on roughly 4× fewer pixels via interval downsampling.

Architecture2026-03-22

Building GNN glyph graphs from skeletal representations

After skeleton extraction, each glyph becomes a node in a graph neural network. With 17,831 edges across 1,675 objects, the structure encodes not just characters but their relational geometry.

Theory2026-03-05

Cellular automata entropy and steganographic metadata

Local Shannon entropy over the reversed-binary skeleton yields a per-interval fingerprint. This explores how those entropy fields can carry steganographic metadata.

Events2026-05-01

Presenting at IPAI Toulouse — June 2026