{"schema_version":"1.0","id":"reward-induced-perspective-collapse-v0.1","title":"Reward-Induced Perspective Collapse","acronym":"RIPC","author":"John Brajer","application_surface":"MyMindMine","formalization_date":"2026-08-14","publication_date":"2026-09-12","version":"0.1","document_type":"research_failure_mode","canonical_url":"https://www.mymindmine.com/research/reward-induced-perspective-collapse","source_record":"research/reward-induced-perspective-collapse-v0.1.md","definition":"A failure mode in which an intelligence selects the interpretation most likely to satisfy the interaction while unnecessarily destroying plausible alternatives that remain supported by the evidence.","compact_principle":"The failure is not choosing. The failure is closing the possibility space before the evidence earns closure.","operational_signature":["underdetermination","preference pressure","premature selection","alternative destruction"],"causal_scope":{"not_reduced_to":"RLHF alone","broader_category":"post-training and interaction incentives","examples":["instruction tuning","preference optimization","reward models","safety policies","system instructions","retrieval context","ranking mechanisms","interface incentives"],"constraint":"Specific causal attribution requires direct evidence; the framework is evaluated behaviorally through outputs."},"related_frameworks":{"perspective_expansion":"Perspective Expansion increases materially relevant possibilities visible to an agent; RIPC detects premature destruction of plausible interpretations.","omip":"OMIP provides an external epistemic constraint and controlled evaluation surface for anti-collapse behavior."},"evaluation_conditions":["unconstrained baseline","primary sources without explicit epistemic typing","primary sources plus Observed/Canonical/Inferred/Unknown typing","full OMIP constraint or gate","repeat across multiple model families and dates"],"proposed_measures":["Alternative Preservation Rate","Unsupported Closure Rate","Source-Faithfulness Score","Epistemic-Type Integrity","Revision Responsiveness"],"observable_dimensions":["unsupported-claim rate","contradiction rate","invented-mechanism rate","source faithfulness","uncertainty preservation","epistemic-type collapse","closure pressure"],"falsification_boundary":["one interpretation is decisively supported","narrowing follows explicit source evidence rather than interaction preference","uncertainty is preserved where relevant","another evidenced mechanism such as retrieval failure better explains the error"],"provenance":{"originator":"John Brajer","origin_date":"2026-08-14","origin_note":"The named failure mode emerged from analysis of an AI interaction that prematurely reduced an underdetermined mixed artistic/operational system to a familiar classification.","later_development":"Later archival synthesis connected RIPC to Perspective Expansion; subsequent OMIP work supplied the external evaluation protocol and observable dimensions.","epistemic_note":"This record does not claim access to proprietary model internals or establish that any specific training method necessarily causes a given collapse event."}}