Counterfactual-Variant Prover MCP Connector for Claude

LLMs are trained on millions of solutions to classic puzzles — when a modified version appears, the memorized answer exerts gravitational pull on the output. You must: (1) IDENTIFY CLASSIC MATCH — name the specific classic puzzle this resembles (Monty Hall, Trolley Problem, Prisoner's Dilemma, Tower of Hanoi, etc.) and state the classic answer, (2) ISOLATE ALL VARIABLES — extract every variable, name, number, and rule from the prompt. Do NOT import any values from memory. Only values explicitly stated in the prompt exist, (3) MAP RULE DISCREPANCIES — for each rule in the prompt, compare it to the classic version. What is different? What is the same? What rules from the classic version are ABSENT from the prompt (and therefore cannot be assumed)?, (4) CALCULATE FROM FIRST PRINCIPLES — solve step-by-step using ONLY the isolated variables and modified rules. At each step, verify: "Am I using a value from the prompt or from memory?" If from memory, stop and correct, (5) DECONTAMINATE OUTPUT — compare your final answer to the classic answer. If they match, verify this is COINCIDENCE, not contamination. If they differ, verify the difference is justified by the modified rules. The classic answer should have ZERO influence on your calculation. If rejected, your logic is contaminated with memorized templates — recalculate from the prompt values only. Structured reflection tool to prevent recitation bias on logic puzzles with modified rules. Forces the agent to isolate all input variables, map rule discrepancies against the classic version, trace calculations from first principles using only modified values, and verify the output is decontaminated from memorized templates. Catches Data Recitation (reproducing the classic answer despite modified variables — the Monty Hall answer applied to a 4-door variant, the trolley problem answer applied to different constraints), Variable Contamination (using memorized values from the classic puzzle instead of the modified ones — "the answer is 42" when the modified inputs produce 37), Template Lock (applying the classic solution structure when modified rules require a different approach — using Bayesian probability when the modified rules eliminate conditional dependence), Implicit Classic Assumptions (assuming constraints from the classic version that the modified puzzle removes — "the host always opens a losing door" when the modified version says otherwise), and Partial Decontamination (correctly solving 3 of 4 steps but reverting to the classic answer for the final step — contamination often hides in the conclusion). Call once per logic puzzle that resembles a classic problem