Auditing and calibrating LLM-driven verification gates

Develop auditing protocols to calibrate large-language-model-driven quality gates used in verification and validation, incorporating adversarial probing, measurement of inter-agent disagreement, and evaluation against held-out physical measurements.

Background

The instrumentation pipelines rely on LLM-driven gates for quality control and verification steps. Ensuring these gates are reliable requires independent calibration and robust auditing procedures.

The authors highlight the need for adversarial testing and disagreement-based diagnostics, anchored by comparison to physical measurements, to validate and monitor these gate mechanisms.

References

Finally, the two annotators are project collaborators rather than independent third parties, and the same pair supplies the judge calibration, so their labels may carry correlated expectation bias. An independent-annotator subset is left to future work.

— GANDR: Claim Auditing for Verifiable Legal Answer Generation  (2609.10293 - Qian et al., 9 Sep 2026) in Appendix A, Section “Human validation of atomic cite_verify”

Nine open questions will determine whether instrumented data matures into a recognised substrate for scientific machine learning. Verification of the verifier. Quality gates are LLM-driven; auditing their calibration requires adversarial probing, inter-agent disagreement, and held-out physical measurements.

— Instrumented data for causal scientific machine learning  (2606.07865 - Wilke, 5 Jun 2026) in Section 7, Methodological questions for the community, Item 3

Evaluation of physical faithfulness is itself an open problem: GenExam's MLLM judge and our binary-checklist evaluator both rely on a vision-LLM to answer the checks, and so remain proxies for expert judgment.

— Towards Physics-Faithful Generation of Scientific Diagrams  (2608.13112 - Zhang et al., 13 Aug 2026) in Section 4.1, “Future Work” (Discussion)

We are explicit about what this paper does not yet establish, in decreasing order of importance. (i) Honest over-rejection is measured on the resolver's own honest generator. Because a well-formed call is defined as one the resolver admits, the $0.00$ over-rejection is a construction-consistency check, not independent evidence; the decisive next experiment is to run the rung against an independent honest-call corpus it did not define (e.g. API-Bank/ToolBench gold calls, or the live models' honest emissions) and report a genuine false-reject rate.

— Closed-World Resolution Against Tool Hallucination in LLM Agents  (2609.19425 - Iyer, 16 Sep 2026) in Section Limitations and Future Work, item (i)

A central open direction is to make learned validators reliable enough to act as control gates, through better calibration, uncertainty over verifier scores, adversarial training against reward hacking, and validators that emit localized, checkable evidence rather than a single scalar~\citep{setlur2024rewarding, kirchner2024prover}.

— Complex Problem Solving in Large Language Models: A Statistical Control Survey and Diagnostic Framework  (2609.20973 - Cai et al., 17 Sep 2026) in Section 5.3, Reliable Validators and Constraint Enforcement

The live arm uses a local service, twelve contracts, and one writer; larger models, distributed services, correlated human decisions, adversarial faults, and longitudinal source drift remain open evaluations.

— Engineering Reliable Commit Gates for Agentic AI: Cost-Aware Verification Portfolios under Common-Mode Data Failures  (2609.10969 - Zheng et al., 10 Sep 2026) in Section 'Engineering Implications and Validity', subsection 'External validity'

Under which conditions such a probabilistic check can still reliably bound the propagation of variation remains an open question.

— LLM Non-Determinism in Deterministic Processes: A Controlled Variation Space for Constraining the Propagation of Non-Determinism  (2610.06255 - Mandl et al., 5 Oct 2026) in Section 8, Conclusion and Future Work

Other open problems include behavior under distribution shift and the choice of audit rate.

— When Can You Ship on Evals Alone? Trial-Level Surrogacy for Offline Evaluation of LLM Systems  (2610.10142 - Schultzberg, 7 Oct 2026) in Section Discussion and Conclusion

Serving the evaluators with a small model is the direction we would pursue next: it addresses the residual external dependency and the dominant cost line at once (Appendix~\ref{app:cost}), and how far judge capability can be reduced before the gate loses the separation of Appendix~\ref{app:separation} is an empirical question we have not answered.

— Verification Turns Capability Gaps into Routing Decisions: Small Language Models for Clinical Documentation  (2610.19169 - Fatehi et al., 13 Oct 2026) in Section 6, Limitations and Future Work