Research foundation
Our consulting work is grounded in original mechanistic interpretability research. The same techniques we publish enable the document classification, extraction, and drift detection we deploy for clients.
Paper I — 2026
Causally Functional Content Representations in Transformer Residual Streams
A literary translation paradigm establishes that content representations peak at ~50% of network depth; subspace activation patching demonstrates single-direction causal effect.
Paper II — 2026
Depth-Dependent Dissociation of Content and Framing in Transformer Residual Streams
Framing representations peak earlier in depth than content, and the two dissociate cleanly. Validated across pharmaceutical, financial, and insurance domains.
In progress
Cross-Architectural Document Classification via Residual Stream Probing
Linear probes on vision-language model residual streams reach up to 100% accuracy on text-vs-chart classification across three architectures. Directly applicable to document routing deployments.