CausalFairAudit: Tamper-Evident Auditing for LLM Fairness Repair

Abstract

Large language models (LLM) are increasingly used to support educational risk prediction, but their decisions may encode sensitive-attribute bias and are difficult to audit after deployment. Prompt-based fairness repair can change model behavior, yet it usually does not record which causal pathways were repaired or whether the repaired decision can be verified later. This paper proposes CausalFairAudit, a tamper-evident audit framework for causal fairness repair in language-model decision systems. The framework converts each prediction and repair action into a canonical audit record containing hashed inputs, hashed model outputs, protected-path evidence, red-line mediators, previous-record hashes, and a Merkle commitment. Using Open University Learning Analytics Dataset (OULAD) based outputs from DeepSeek-V4 and Qwen3.6, we compare an Original Agent, Prompt-only Repair, and CausalFairAgent. Results show that prompt-only repair may improve accuracy while worsening causal fairness or suppressing high-risk predictions, whereas causal repair performs localized intervention on identified unfair pathways while preserving task-relevant evidence. The audit ledger generates 6,000 records at 12,981.1 records/s and detects 100% of simulated single-record tampering.