At a Glance
Keep a human in the loop and record the evidence and edits: letting stewards challenge agent reasoning before commit prevents single AI mistakes from becoming system-wide errors and makes every change auditable.
ON THIS PAGE
What They Found
Agents can propose metadata using cited documents and structured fields, but automatic commits risk spreading mistakes across systems. A contestable workflow where multiple specialist agents submit supporting and opposing arguments, and a human steward can accept, reject, or modify those arguments before commit, preserves the full reasoning chain. Storing an append-only decision record with agent identities, cited evidence, and steward interventions makes decisions reconstructable and lets future audits or re-scoring reproduce outcomes without rerunning models. contestable workflow and the ability to submit supporting and opposing arguments are central to this approach; the system can also rely on an append-only governance record to keep a verifiable trail.
Key Data
1In the synthetic demo of 247 telemetry points, 229 (≈93%) would receive a short, evidence-based steward review while 18 (≈7%) required full argumentation and deeper review.
2Before human intervention in the example case, agent argument margins were nearly tied: 0.045 vs 0.044 for the two leading labels, showing how small score differences can hide real ambiguity.
3A steward’s edits increased the decisive margin to 1.758, producing a committed classification; normalized scores (1.00 vs 0.99) could have misleadingly suggested high confidence without raw-margin context.
Implications
Data stewards and governance leads: to design workflows that prevent AI errors from becoming permanent catalog entries. Engineers building agent-assisted metadata pipelines: to implement triage, evidence citation, and append-only provenance so agent outputs are contestable and auditable. Technical leaders evaluating agent reliability: to balance automation speed with human oversight and measurable reviewer interventions. For teams, see Multi-Agent Knowledge Management use case and Blackboard Pattern for coordination.
Need expert guidance?We can help implement this
Key Figures

Fig 1: Figure 1: Contestable governance for HVAC point classification: data analysis, multi-agent argumentation between two agents, human contestation, and selection and storage of a verified onboarding record.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
The reference implementation is a synthetic demo and does not measure its effect on real-world accuracy, workload, or error rates. The scoring method shown is illustrative and uncalibrated; base scores can vary between agent runs and correlated arguments can inflate margins. Key features—an append-only promotion gate, triage heuristics, and writing verified precedents back into the system—remain proposed and need production validation. See Evaluation-Driven Development (EDDOps) for how such features might be validated in production.
Methodology & More
The proposed workflow splits metadata onboarding into four clear phases: (1) data analysis that gathers structured fields, relevant documents, and similar past decisions; (2) multi-agent argumentation where specialist agents submit targeted supporting or attacking arguments with cited evidence multi-agent argumentation; (3) human contestation where a steward can accept, reject, modify, or add arguments and see scores recomputed live; and (4) selection and storage where the chosen class is explained and an append-only governance record is written. The record attributes every decision to the agents and named reviewers, cites the evidence used, preserves rejected reasoning, and makes the full lineage retrievable in one query so an auditor can recompute outcomes without re-running models. An open reference implementation illustrates the pattern on HVAC point classification: of 247 synthetic points, 229 would get a shorter review while 18 receive full argumentation. One detailed scenario shows two agents tied on a point label until a steward rejects a name-based argument, adds a document citation, and raises the margin decisively, after which the system commits the corrected class and stores the full contestation history. The paper emphasizes that such records improve traceability and downstream safety but do not guarantee correctness: scoring rules need calibration, versioned glossaries and promotion gates must be implemented, and empirical studies must measure effects on review time, error rates, and whether verified precedents improve later accuracy. Event-Driven Agent Pattern and Mutual Verification Pattern provide complementary perspectives on how to structure the phases and governance around agent outputs.
Need expert guidance?We can help implement this
Credibility Assessment:
Authors have modest h-indices (6–8) but no affiliations or venue prestige; arXiv preprint and no citations suggest limited established credibility.