Evidence, citations, and the right to abstain

What it is, and what it is not

A citation in Manufacturing PLM is a reference to a specific object at a specific revision, attached to a specific claim in an answer. Clicking it opens that object as it was read, and re-running the underlying query returns the same result.

It is not a footnote listing documents the model looked at. That pattern is common and close to useless: it tells you the search retrieved something without telling you whether the sentence above it follows from what was retrieved.

It is also not a confidence score. A number between zero and one attached to a claim gives a reader something to feel rather than something to check, and in an engineering organisation the difference is the whole point.

The mechanism: revision-addressed retrieval

Every embedding carries the revision of the object it was made from, and this could not have been added later. An embedding with no revision on it cannot be distinguished from a current one, so the first time a part revises, semantic search begins answering about a design that no longer exists — confidently, with nothing to notice it by.

When a revision supersedes, its embeddings are invalidated rather than left to age. Retrieval therefore returns content that is either current or explicitly historical, and the citation says which.

Underneath, agents call the same domain services the interface calls — the same resolver, the same rollups, the same permission checks. That is what makes a citation reproducible: the reader clicks through to a resolution they can run themselves, under the same conditions, and compare byte for byte. An agent with a private read path could not offer that, however good its answers were.

The failure it prevents

An engineer asks why a product's cost moved. The assistant returns a fluent paragraph naming a component and a supplier price increase. It is specific, plausible, and wrong — the component it names was superseded in July and the price it refers to was read from a cached figure eleven months old.

Nothing about the answer looks wrong. There is no phrasing that signals a stale retrieval, and the engineer has no way to check it short of redoing the analysis, which is the work they were trying to avoid.

They act on it, or they check it and stop trusting the feature. Both outcomes are worse than no feature at all, because both consumed the reviewer's attention. Citations that resolve to a revision make the check cheap enough to actually perform, which is the only thing that makes an AI answer usable in a regulated conversation.

How it meets the rest of the product

Deterministic services underpin the citable claims. Impact analysis, resolution and comparison have no model in their path, so the same question asked twice returns identical bytes — which is precisely what makes their output quotable. The agent's contribution is framing and judgement layered on top of results a person can independently reproduce.

Abstention is a first-class outcome rather than a failure mode. An agent that cannot support a claim with evidence says so and escalates, and there are object classes agents may not act on at all regardless of permission tier.

Every answer lands in the hash-chained audit trail with the agent named, the tier it ran at, the tools it called and the revisions it cited. Six months later, “where did that number come from” has an answer that names a system, a run and a set of revisions rather than trailing off into a shrug.

Citations also inherit permissions. An agent cannot cite an object the asking user may not see, and it does not reveal existence through a refusal — the answer is shaped as though the object is not there, because for that user it is not.

How it meets your ERP

Values read from the ERP are cited differently from values Manufacturing PLM owns, and visibly so. A cost figure carries the source system and an as-of timestamp, because it is a cached read of somebody else's master record rather than a fact Manufacturing PLM can vouch for.

That distinction matters in exactly the analyses people most want to automate. When the Cost Agent explains a fourteen percent increase, the structural half of the answer is reproducible from Manufacturing PLM's own resolution; the price half is a read of your ERP at a moment in time, and conflating the two would make the whole answer look more solid than it is.

Agents also inherit the system-of-record matrix. Because cost is ERP-mastered, the Cost Agent proposes a structural change rather than a cost edit — and cites the lines that drive the number, at the revisions it read them at.

Where the boundary is

A citation proves what the agent read. It does not prove the reasoning between the evidence and the claim is sound. A well-cited answer can still draw the wrong conclusion, which is why the permission tiers exist and why most agents sit below the level at which they can act unilaterally.

Citations also cannot reach outside the tenant. Component data feeds, supplier declarations and CAD metadata are cited to the record Manufacturing PLM holds and the date it ingested, not to the upstream vendor's live database — an honest description of what was actually read.