praxagent / Technical record
Research Notes
Experiments, methods, controls, and engineering notes from work on model internals and AI systems.
A Linear Nudge, a Nonlinear Wake
We followed controlled hidden-state edits through the final 29 blocks of Llama 3.3 70B. The delivered edits were dose-linear and a fixed corpus-average Jacobian preserved that scaling, while the model’s actual downstream response did not.
Why Experimental Details Matter: Recovering from a Flawed Interpretability Study
A striking interpretability result survived preregistration and still came from a flawed comparison. We trace how prompt echo, a domain-mismatched lexicon, and unequal stakes produced an overconfident story; retract the …
Can a Jacobian Lens Detect SAE Steering?
A prospectively frozen Llama 3.3 70B experiment asks whether SAE steering leaves a detectable downstream fingerprint in Jacobian-lens space. A preregistered follow-up adds semantic hard negatives, same-subfamily …
Open-sourcing (and Auditing) a Jacobian Lens for Qwen3.5-397B
As of July 10, 2026, Praxagent is open-sourcing the public Jacobian lens fitted for the largest base model so far: Qwen3.5-397B-A17B, with a pre-registered readout audit against identity and random-J controls. Weights, …
How to Read an SAE Feature ID
A primer on sparse autoencoders: what a feature ID is, how labels get assigned, and why an activation map is not yet an explanation. A public deception/roleplay feature set is used as a worked example under Llama 3.3 70B …