mechinterp

A study of introspective access under persona induction using Activation Oracles, verbal self-report, and activation-space similarity on Qwen3-8B.

Which layer is your Jacobian lens fitted to? The library says final, one line of the paper says penultimate — with dependency-free checks to tell them apart.

An interactive primer on mechanistic interpretability of LLMs

Full pipeline, protocol and results for "Read, Replace, Never Rewrite: Continuous Thoughts Carry Search Structure, Not Answer Identity", the first causal test of the COCONUT superposition hypothesis.