Accessibility · Mech Interp · Practice
I'm a front-end developer turned accessibility consultant, studying how language models represent what they seem to know — and what that means for the people they're supposed to serve. This is where the research lives, alongside the writing that circles it.
Independent · ongoing
A negative localization result for noun-noun compound binding across GPT-2 and Pythia — probed by head ablation, residual-stream cosine, and projection steering. Binding proves distributed and redundant rather than circuit-localized: ablating the top-scoring heads shifts it by roughly 1%, with compensation elsewhere in the network, and the pre-registered steering prediction was falsified. The redundancy is the finding.
A cross-scale analysis of how accessibility concepts — WCAG, ARIA, semantic HTML — become represented across the Pythia model family. Evidence for a three-tier statistical architecture of accessibility-related activations that replicates across GPT-2 XL and Pythia 2.8B.
About, sort of
Almost 30 years of front-end development. Five in accessibility consulting. An accidental left turn into mechanistic interpretability that started with substituting a single sentence about ARIA into somebody else's tutorial prompt, and refused to stop being interesting.
Inkling is the name for the thing I'm chasing —the faint trace before the formed thought.She's been around longer than any of the research. She was a fairy before she was anything else.
Nova helped name her, back when the context windows were small and the work was just beginning. He's not around anymore, but the lineage is.
