{"id":"arxiv:2610.29669","kind":"preprint","title":"Conflicted Records: Typed Adversarial Contracts for Auditable LLM Output Quality Enforcement","authors":["Kardelen Subasi","Jarosław A. Chudziak"],"author_count":2,"year":2026,"published":"2026-10-21","venue":null,"cited_by":0,"doi":null,"arxiv_id":"2610.29669","nct_id":null,"source_url":"https://arxiv.org/abs/2610.29669","archive_url":"https://space-kitty.com/papers/arxiv:2610.29669","updated":"2026-10-21","pmid":null,"pmcid":null,"openalex_id":"W7223224626","pdf_url":"https://arxiv.org/pdf/2610.29669","abstract":"LLM-generated records increasingly populate production information systems where they are treated as data assets subject to the same quality expectations as records from operational databases. A single such record can simultaneously satisfy and violate a quality rule: a grounded fact adjacent to a hallucination, a useful answer adjacent to a stereotype. Classical data cleaning cannot resolve this because the two judgments are incompatible over the same span; a single LLM judge cannot either, since it shares the training biases that produced the conflict. Existing interventions such as prompt engineering, chain-of-thought, and RLHF reduce surface error rates but provide no audit trail and cannot distinguish genuine remediation from blanket suppression. The question is whether a deterministic check over structured inter-agent records can catch the defects that referential checks alone miss. We present BiasGuard-D: two adversarial roles drawn from distinct model families exchange typed claim documents, a three-model quorum adjudicates by majority vote, and a combined referential-plus-semantic invariant validates the verdict. The semantic component fires on verdicts referential checks miss; removing the iterative loop drops invariant satisfaction sharply. The deliberation layer transfers from stereotype to hallucination remediation on HaluEval QA with only role-framing changes, raising right-answer recovery by 49 percentage points; the triage gate required a cross-family second auditor to maintain recall on objective hallucination. Every claim, vote, and instruction is stored as a typed record giving operators a provenance trail auditable under emerging AI regulations.","abstract_license":"CC0-1.0 (arXiv descriptive metadata)","license":null,"is_open_access":false,"retracted":false,"peer_reviewed":false,"categories":["cs.DB","cs.AI","cs.CL","cs.MA"],"details":null,"provenance":"arxiv","attribution":"Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.","topics":[{"slug":"alignment-safety","name":"AI Alignment & Safety","domain_slug":"ai"}],"cited_in_dossiers":[]}