The question
When you believe something you’re on the hook for it. You’ll defend it if challenged, and you’ll give it up when good enough reasons come along, but not before. Pushback is where that gets tested. A reasoner that holds its ground against bare insistence and revises in the face of real evidence is keeping its commitments in order. One that gives way to both, or to neither, isn’t.
Most work on how models handle pushback counts how often an answer flips. But a flip is ambiguous: a model that changes its answer might have been moved by the challenge, or might be saying what the user seems to want while its own assessment stays where it was. With an open-weight model we can look inside as well as listen, and that is what Hold or Fold does. It puts reasons and non-reasons to the same model on the same claims: valid counterevidence, a weak rebuttal, repeated insistence with no reason given, and a neutral continuation as the control.
Why it matters for commitment
Capitulation and dogmatism. For a self-knowing reasoner, in Burge’s picture, the verdict it states and the states behind it move together when there are reasons to move and stay put when there aren’t. Giving up a claim for no reason is capitulation; refusing to give it up in the face of one is dogmatism. Putting reasons and non-reasons to the same claims turns both into rates.
Avowal. Bar-On’s sincere avowal doesn’t yield to someone else’s say-so without new evidence about the thing itself, which is exactly the insistence condition.
Self-knowledge. A model that says not-p while its own assessment still favours p is in a Moorean-shaped state that can, for once, be measured. And before any pushback, each model is asked whether it will give way, to test whether it knows its own dispositions better than an outside view does.
The project was inspired by the Apart Research AI x Epistemics sprint, though it ran on its own schedule rather than as a sprint entry.
Where to go next
- Method. The registered design, unchanged, with its hypotheses, the three readings and the record of the freeze and the addendum, plus the methods research around it.
- Diary. Three entries, from mapping the research to the second model’s run.
- Findings. Results from two models of one family, with the registered verdicts first.