Contradiction Lift
Contradiction Lift Skill
Have Claude and Codex independently solve the same problem, surface the contradiction from where their answers diverge, and pursue Aufhebung (sublation / 止揚) — not averaging, not compromise, but a lift that preserves both truth-moments while raising them to a higher frame. The output is a selection mechanism (when does A win, when does B win, and why) or an honest aporia (this cannot be lifted; here is the irreducible axis).
Overview
The single most important design fact: the enemy of Aufhebung is not conflict — it is averaging (a watered-down middle ground). Left alone, two LLMs fail in two ways: premature agreement (mutually sycophantic "great point, I agree" → instant collapse), or sterile gridlock (debate where each just re-states its position). Telling them to "find the middle" produces the worst outcome — a diluted average. So almost the entire mechanism is spent keeping the process out of averaging and out of gridlock, and forcing a real lift.
A Contradiction Lift run:
- Fixes the question (Decision Contract) so interpretation gaps are not mistaken for contradictions.
- Seals two independent solutions — a fresh Claude subagent (Solver A) and Codex (Solver B) solve in parallel; the orchestrator dispatches both and reads neither until both return, so sealing is structural (the orchestrator never authors a solution it could contaminate).
- Maps the divergence by type and names the load-bearing premises (flip-test).
- Routes each disagreement: empirically decidable → Codex runs a discriminating experiment; the rest → preservation.
- Preserves each side via mutual steelman (accept / repair-once handshake).
- Constructs a lift — a selection mechanism
f(C) → A | B | N, not a position. - Audits the lift against 7 tests; if it fails twice, declares an honest aporia.