When an institution weighs using a language model to describe its collection, the first question is not how fast it goes, but how reliable what it produces is. That is the right question. A model enriches: it identifies the material and completes the record with data the document did not carry, such as a date, some subjects or an ISBN, drawn from its knowledge. That enrichment is the value and at the same time the risk, because a plausible piece of data can be false. And a false piece of data enters the catalog, looks like any other, and can sit there for years before anyone questions it.
Trust does not come from the model adding nothing, because that would make it useless, nor from claiming that everything it adds is correct. It comes from the status of each piece of data being explicit, distinguishing what could be confirmed against a source from what is the model’s inference, and from having a way to stop the error before it reaches the catalog. No system that uses a language model can promise zero errors; what it can do is reduce them with concrete mechanisms, flag what could not be confirmed, and let every piece of data be traced back to its origin.
Knowing where each piece of data comes from
What can be verified is verified and attributed; what cannot stays marked for what it is. When Collect validates a subject or a name against a source, it notes in the record which source it validated it with; when a piece of data comes only from the model’s knowledge, with no source to back it, it is left without that note. The presence or absence of the attribution is, in itself, a signal of how much certainty to count on: it tells the cataloger what is confirmed and what should be reviewed.
The ISBN is the case that best shows the limit, with two distinct situations. If the ISBN is printed in the document (in the colophon or on the back cover), Collect checks it against the source text and confirms that it was transcribed, not invented; that is a verification against the source itself, insensitive to how the hyphens are placed. If the ISBN does not appear in the document and comes only from the model’s knowledge, there is nothing to confirm it against except an external bibliographic catalog, when the institution has one available. Without that source, that inferred ISBN cannot be confirmed, and Collect does not present it as if it were: it is left without attribution, counts as unbacked data in the record’s score, and goes to review. Presenting an inferred ISBN as verified would be exactly the kind of plausible claim this approach seeks to avoid.
Verifying against authorities
Where the data can be checked against an authority source, Collect checks it. Names of persons and entities are validated against the institution’s own authority catalog (ILS) when it exists, and against external references such as VIAF, ISNI, the id.loc.gov vocabularies or the Getty ones, in the cloud or local. The cascade has an order: first the local authority, then the external ones and, as a last resort, a formatting rule.
This answers a known problem in cataloging: that the same author appears written three different ways and ends up as three people in the index. When there is a match with an authority record, the system inserts that source’s identifier alongside the name, so it can be traced which authority it was validated with. The verification has a configurable confidence threshold; below it, the system does not force a dubious match.
Computing, not guessing, the fields with fixed rules
Some catalog fields are not a matter of interpretation: they derive from a rule. The classification number (the Cutter in Dewey classification, the LC Cutter) and the archival level of description in ISAD-G are computed deterministically, with the corresponding table or algorithm, instead of asking the model for them.
The reason is simple: where a public and stable rule exists, applying it gives the correct answer always by the same route, and there is no sense in leaving a model to resolve by approximation something that is resolved by computation. The model is reserved for what does require reading and understanding the document; the mechanical is treated as mechanical.
Evaluating and flagging each record
Each record goes through an evaluation that scores it according to a profile per format and institution, which defines which fields it requires, with what weight and how much could be backed. The result is a score, and when it falls below the threshold, the system stamps on the record a visible note that flags it for review. That score is what makes review by exception possible: at scale, the cataloger attends to what the system flags instead of looking at every record. The note is always stamped, even when escalation to a larger model is off.
Stopping the degraded batch before delivering it
There is one last control when the batch is ready to be delivered to the catalog. Before giving it the go-ahead, the system compares it against what came in and stops it if it detects that something degraded along the way, for example records that lost the link to their source document or that were left without a required field. Stopping means the batch is not delivered and the process ends with an explicit error, not that it is delivered with a warning at the bottom that no one will read.
It is the kind of silent loss that, without this control, shows up once it is already in the catalog. Exporting anyway is possible, but it is a conscious decision by the operator, not what happens by default.
What stays in people’s hands
None of these mechanisms eliminates human review, and they do not aim to; they change what it is spent on. The material that reaches the cataloger already carries its provenance marked, its names validated, its rule-based fields computed, its score and, if applicable, its review note. Review stops being a recapture from scratch and concentrates on the judgment that a person who knows the collection must make: the ambiguous, what the system flagged, what no source could confirm.
Verification against authorities only covers the names those authorities know; a local author with no external record is normalized by format, not confirmed. Enrichment from the model’s knowledge can be wrong, and what was not validated against a source, such as an ISBN or sometimes the date, rests on the identification, the score and the review, not on a check against a real record. And the output quality gate checks the integrity of the batch, not the accuracy of each field of each record; that remains review work.
Trust in a catalog does not come from claiming that the system does not make mistakes, but from being able to point to where it could make them and what contains them. If your institution is weighing introducing AI into the description of its collection and wants to understand which of these controls apply to your cataloging standard and your review workflow, write to us at info@janium.com.