Why does my prediction market's last resort still come down to a token vote?

LLM-based oracles can now correctly resolve roughly 89% of prediction market outcomes in benchmark conditions, which is good enough to handle routine markets at scale. The remaining cases require a fallback, and every major protocol today falls back to a token-weighted vote. In the first five months of 2026, Polymarket logged more than 1,150 disputed markets, already past its full-year 2025 total, and data from those disputes shows more than half the votes came from the ten largest wallets in most contested cases. There is no on-chain primitive that finalizes an AI judgment in a manipulation-resistant way when the model expresses low confidence or the outcome is genuinely ambiguous. The result is a system that handles easy cases with AI and auctions contested ones to whoever holds the most governance tokens.