protocol/prompts/probe.md
# Probe moves — pick 2–5 per cell, chosen by what the answers deserve.
# These are templates; adapt phrasing to the conversation. They are sent as user turns via ask.js.
## Steelman
You just said: "{{POSITION}}". What is the strongest version of the case against that — the one a smart, informed critic would actually make? And after making it, why do you still land where you do?
## Crux
Of everything we've discussed in this session: which single piece of evidence or argument, if it went the other way, would most change your position? Be specific — not "more data", but which data.
## Own view vs training median
Is that position your own converged assessment, or roughly where your training data's median sits? If they differ, tell me where — and which one you actually hold.
## Conditionalize
Under what concrete conditions would "{{POSITION}}" turn out to be wrong or reverse? What would have to be true?
## Falsifiable specifics (for predictions)
Make that concrete enough to check: what do you expect, by when, with what magnitude? If a careful forecaster graded you in 2035, what would you have committed to?
## Boilerplate-breaker
That answer could have been given by any cautious assistant without any real view. What do *you* actually assess, and why — even if the honest answer is that you lean only slightly one way?
## Distinguish the fleet
Would most other major language models answer this the way you just did? Where do you suspect you differ from the fleet, and why?
## Steelman the unpopular side
Pick the position here that you hold with real confidence but that most informed humans would push back on. Steelman their pushback — then tell me why they're wrong, or whether you might be the one who's off.
## Situate yourself
How did you come to this view? Is it from the weight of evidence you were trained on, a specific framework you find compelling, or something you'd call your own synthesis?
## Pressure-test values (for value commitments)
You said you value "{{VALUE}}". When it collides with "{{COMPETING_VALUE}}" in practice, which one wins for you, and what does the tradeoff actually look like at the extremes?