Inference Battery Token (IBT): Failure Compensation as Training Signal

Abstract

This paper specifies the Inference Battery Token (IBT), an ERC-20 on Base that mints when the inference orchestrator verifies a failed AI response, compensating the subscriber and capturing a structured DPO preference pair in one operation. Minting occurs through two pathways: automatic spot-checks that re-execute sampled requests on trusted nodes, and subscriber-initiated flags confirmed through the same re-execution pipeline. Each verified failure mints tokens proportional to the compute cost wasted, ranging from 1 IBT for embeddings to 20 IBT for extended reasoning. Subscribers redeem IBT for battery credits, permanently burning tokens. A genesis allocation of 10,000,000 IBT bootstraps exchange liquidity, network operations, community grants, and early investor capital with 12-month vesting. Because minting depends on AI failure rates and redemption increases with adoption, circulating supply contracts as models improve. Contributors who serve compute earn platform credits and elect payouts in USD or IBT, with a 15% bonus for token election. Hardware contributors may also earn IBT by generating zero knowledge proofs for the verification system during idle GPU windows. Revenue from selling accumulated failure records to frontier labs through the data access tiers described in the companion paper provides fundamental backing for token value.