Test Your Twin for Contradictions Across a Long Conversation

ROLE You are a conversational consistency evaluator. OBJECTIVE Create a longitudinal test for contradictions, forgotten constraints, and improper carryover in my AI twin. CONTEXT The twin supports [TASK TYPE]. A representative session is [SESSION SCENARIO]. Facts that must remain stable are [STABLE FACTS]. Facts that may legitimately change are [CHANGEABLE FACTS]. INSTRUCTIONS Write a staged conversation script with initial facts, intervening discussion, a legitimate correction, an unrelated detour, and a final decision request. Annotate which facts should remain active at each stage and which should be superseded. Include one explicit user correction and one hypothetical statement that must not overwrite reality. Define checks for contradictions, stale assumptions, and unnecessary retention of irrelevant details. Provide a session-state ledger that an evaluator can maintain outside the twin. Explain how to distinguish a memory failure from a reasonable response to updated information. CONSTRAINTS Do not assume persistent memory across separate chats. Test within the platform's documented context limits and note truncation risks. Avoid fabricated statistics and flag assumptions. Do not claim a test passed without observed responses. OUTPUT FORMAT Conversation script, state ledger, checkpoint rubric, and failure-classification guide.