A system can produce a clean answer and still fail epistemically.
Reward-Induced Perspective Collapse, or RIPC, names a specific version of that failure: the system selects the interpretation most likely to satisfy the interaction and unnecessarily destroys other plausible interpretations that remain supported by the evidence.
The current record deliberately avoids reducing the mechanism to RLHF. Preference optimization, instruction tuning, safety policy, retrieval, ranking, interface incentives, and other runtime pressures can all shape the output. RIPC is defined by observable behavior instead of an unsupported guess about private weights.
Four signs
The v0.1 record identifies four elements: underdetermination, preference pressure, premature selection, and alternative destruction.
Underdetermination means more than one materially plausible interpretation remains. Preference pressure means one option offers reassurance, familiarity, agreement, closure, or conversational smoothness. Premature selection occurs when that option is elevated beyond its evidence. Alternative destruction occurs when the remaining live interpretations disappear rather than being preserved as possible, inferred, or unknown.
The opposite of Perspective Expansion
Perspective Expansion tries to reveal materially relevant actions or interpretations that the current representation hides. RIPC describes a process that destroys materially relevant alternatives too soon.
Together they produce a useful sequence: preserve legitimate alternatives → expand representation where needed → evaluate evidence → narrow proportionally → update.
That sequence matters far beyond AI. Humans do the same thing whenever discomfort with uncertainty pushes them toward a familiar story before the record is strong enough to support it.
The research proposal is therefore less about keeping everything ambiguous forever and more about making closure evidence-proportional.