Your agents write metadata into your systems — how do you know which human actually stands behind it?
5 min read
In most organizations today, the honest answer is that you can't. Agents built on large language models can propose metadata, cite documents, and flag quality issues, and they are being pointed at exactly the fields where a wrong value travels far: telemetry tags, point classifications, transaction attributes. Humza Akhtar and Brian Freeman, in a paper submitted to arXiv on 22 October 2026, locate the weak spot not in the proposing but in the approving. Unreviewed proposals can spread errors into downstream systems, and a perfunctory approval leaves no record of what a reviewer checked. Once that happens, the human behind a committed value has vanished from the record, and the audit question — who stands behind this entry, and what did they actually look at? — has no answer.
Their proposed fix, which they call contestable governance, starts from that observation and makes the steward's intervention the record. Specialist agents produce scored, cited arguments for disputed metadata decisions. A human steward then accepts, rejects, modifies, or adds arguments before anything is committed. A governance record retains the evidence, the interventions, and the outcome. So when you later ask who stands behind a value, the answer is not "someone approved it." It is a specific person, a specific action — accepted, rejected, modified, or added — with the arguments and citations that were on the table at the moment of the decision.
Three supporting pieces keep the pattern usable. A triage path reserves full argumentation for uncertain proposals while preserving review before commit, so the heavyweight process goes to the decisions that need it rather than stalling the routine ones. A versioned domain glossary supplies model-specific terms and acronym definitions to the agents. And a gate controls which verified decisions become precedent, so one confirmed case does not quietly bind every future one. The paper, at 15 pages with two figures and two tables, describes the full workflow and the scoring and record models; the account here follows what the abstract establishes.
The authors illustrate the pattern with two HVAC use cases. The first is point classification using Brick Schema class labels, supported by an open reference implementation. The second is an illustrative warranty failure-mode decision linking claim fields and telemetry to service notes and revisioned bulletins — the kind of judgment that in many buildings currently rests with a small group of experienced people. That second example is explicitly illustrative: it shows the shape of a contested decision, and it establishes no observed product result. The arXiv listing links two ancillary files alongside the paper — reproduce_warranty_scores.py and warranty_score_results.txt — whose names indicate they are there to let a reader recompute the warranty example's scores instead of taking them on faith.
Where this lands for the audit question is a shift in what gets held accountable. Much of the attention on agent oversight goes to watching the agent itself — its prompts, its tool calls, its outputs — and catching misbehavior before it commits. Contestable governance takes a different position: assume the agent will propose things, some of them wrong, and put the accountability in the record instead. The design's promise is traceability, not infallibility. It does not claim that every committed value is correct; it claims that each one carries a named human intervention and the evidence that was weighed. Whether that record satisfies a particular auditor or regulator is for the organization to judge; the abstract points to the scoring and record models as what to measure against.
If you want a working test for your own agent-written fields, one adaptation of the pattern — offered as an editorial suggestion, not a finding measured in the paper — is to run three questions against any field an agent has filled. Can I name the recorded human intervention behind this value? Do the arguments and citations the decision rested on still exist, retrievable, in the record? And is this decision treated as precedent, or as a one-off that the next similar case can contest? If the first question has no answer, the field is carrying an invisible author.
The same discipline applies one level down, to the notes you keep about your own checking. SentX SI's public model card advises keeping a portable checkpoint of checked findings and open questions for work across conversations, and it pairs that advice with a candid caveat: the checkpoint builder formats your notes; it does not verify them. The card's known-limits section makes the same point from the other side — answers can contain incorrect facts, citations or calculations, so open cited sources and check the supporting passage. A formatted checkpoint is worth exactly what the checking behind it was worth, no more, and the value of the record lies in the retained evidence and interventions, not in the formatting. One boundary to keep in mind if you submit material there: under SentX's privacy policy, submissions are not confidential and may enter training or shared memory, so handle sensitive operational data accordingly.
Source: Victoria AI Model Card | Capabilities, Evidence and Limits | SentX SI; Privacy Policy | SentX SI (sentx.ai)
The deeper bet in Akhtar and Freeman's design is that perfunctory approval is not a minor logging gap but the mechanism by which accountability evaporates. Make the intervention visible, scored, and retained, and the record begins doing the work the approver stopped doing. For teams whose agents are already writing into production systems, that is the difference between a metadata field and a piece of evidence.
Sources
- Contestable Agents for Data Governance: Recording Human Stewardship of AI-Generated Metadata Across Structured and Unstructured Systems — arXiv
Supported claims and limits
- Paper arXiv:2610.31553 by Humza Akhtar and Brian Freeman, submitted 22 Oct 2026; 15 pages, 2 figures, 2 tables
- Abstract: LLM agents can propose metadata, cite documents, flag quality issues; unreviewed proposals can spread errors downstream; perfunctory approval leaves no record of what a reviewer checked
- Contestable governance: specialist agents produce scored, cited arguments for disputed metadata decisions; a steward accepts, rejects, modifies, or adds arguments before commit; a governance record retains the evidence, interventions, and outcome
- Triage path reserves full argumentation for uncertain proposals while preserving review before commit; versioned domain glossary supplies model-specific terms and acronym definitions; gate controls which verified decisions become precedent
- Two HVAC use cases: point classification using Brick Schema class labels supported by an open reference implementation, and an illustrative warranty failure-mode decision linking claim fields and telemetry to service notes and revisioned bulletins
- arXiv page links ancillary files reproduce_warranty_scores.py and warranty_score_results.txt; their contents were not opened, so only their presence and names are established
- Victoria AI Model Card | Capabilities, Evidence and Limits | SentX SI — SentX SI
Supported claims and limits
- First-party product disclosure, not independent certification
- Intended uses advise keeping a portable checkpoint of checked findings and open questions for work across conversations
- Checkpoint builder formats your notes; it does not verify them
- Known limits: answers can contain incorrect facts, citations or calculations; open cited sources and check the supporting passage
- Privacy Policy | SentX SI (sentx.ai) — SentX SI
Supported claims and limits
- Submissions to SentX are not confidential and may enter training or shared memory; handle sensitive operational data accordingly