Emotion Axes in a Coding Agent

What I value most here is that this validates an internal metric for welfare measurement. Getting an activation-based affect readout that actually discriminates conditions in an agentic setting, with proper controls, is something prior welfare work, including work I was involved in, was unable to get working, and this project does it convincingly. The experimental controls and validation are unusually strong: the fixed neutral readout suffix so the tokens at the readout position are identical across conditions, the readout layer chosen without touching session data, the comparison against self-report and text baselines, the phase-matched test ruling out rigged sessions simply running longer, and the cross-model observer analysis. The main remaining question for me is generalization: whether the same relationships hold across more model families, task domains, and elicitation settings, and the observer result itself shows what one model represents doesn't automatically transfer. Overall, I found the results clear and convincing within the stated scope.

In testing whether emotion vectors are a good measure of possible model welfare states, the project does well to focus on stimuli that are plausible welfare candidates given how models are designed: approval by a human user and ability to complete a task. The results are patchy and hard to interpret. A similar methodology with bigger models and more extensive work to identify and vet the candidate emotion vectors could be fruitful.