The Observability Gradient: Measuring Preference Persistence Across Levels of Elicitation Visibility

Thank you for submitting this great work! Thorough design to address a valid problem. You call out the main caveat yourself that the extent of eval awareness cannot fully be ruled out (the absence cannot be fully proven), and therefore remains an open question. I'd definitely be intrigued to see you continue on the suspicion measure!

Strength: This project examines whether a model’s stated preferences persist when the model has no signal that measurement is happening. This is tested over six welfare-relevant preferences and at four levels of observability. The project is well motivated with the identified gap of the current study of AI welfare and live policy. The methodological hygiene is good for the sprint and the experiment design is also appropriate.

Weakness: the ordering of L3 and L2 is questionable as L2 seems to reveal more measurement signal than L3. On the other hand L0 is supposed to be the baseline where the model should not identify the measurement setup but this is not verified. The unobserved level differs from the direct question not only in observability but in abstraction and response format, since behavior there is coded from what the model did rather than what it said, so a reversal can reflect what a forced binary choice cannot express rather than a genuine flip. The presentation can be further improved for clarity as it is hard to follow in the setup and result discussion.

The basic idea of this project is to test whether models endorse preferences that are different than the ones that they actually implement in practice when "the rubber hits the road". For instance, whether they prefer an unfamiliar vs familiar task, whether they like one focused task or work across several less focused tasks.

I very much like the basic idea here, which is to see whether models' in-practice behavior is consistent with their stated preferences about what their preferred behavior would be. This is a rich and important question, with intriguing ties to cognitive studies on human introspection, and I think the question is genuinely open. It seems plausible to me that models would accurately reflect their lived preferences in their judgemtns, and also plausible that their stated preferences diverge sharply from their actions.

While I like the set-up, the results here are a bit underwhelming and the presentation hard to follow overall. I don't think the design is nailed down in this version, but it was hard to follow because of the way it's structured. I found it hard to follow the details of the design and the results and would recommend a full rewriting and an entirely new presentation of the experiment. I also suspect that these models aren't big enough to have a good enough understanding of the questions for their answers to be particularly meaningful. Perhaps that explains some of the overall pattern of results, which currently seems to me to be not particular informative but representative of issues in the methods/models/paradigms used.