Make the path from source to statement visible

I examine the transformations between the source image and the text a user searches, including reading order and the association of a passage with its page. In a deliberately mixed review of 520 pages, fifty-two showed ordering problems that a simple file-count check would not have revealed. The review informed a correction programme, but it was not a random sample supporting a population-wide error estimate. I want the product account to distinguish those purposes. Collection owners and specialist reviewers help determine what the source means; the data team makes the transformation and its limits visible. Corrected text should remain traceable to the source, with a way for a user to inspect the page rather than assume that a convenient search extract is an authoritative replacement for the original.

Two descriptions can resemble one another because they concern the same record, different copies, related items or a repeated form. I use matching as a way to identify candidate relationships, not as authority to delete one side. In our review, eight hundred candidate pairs led to 210 accepted duplicate pairs and 590 retained as distinct after specialist examination. Those are pair-level decisions, not a count of records destroyed or a general accuracy score. The collection's identifiers and provenance remained available. I ask who is authorised to decide the relationship and how the product preserves a later correction. A useful linking method should make the catalogue easier to navigate without making its history impossible to reconstruct when a confident-looking match turns out to have been too simple.

An exact date, a year and a range contain different kinds of knowledge. I want that distinction represented in the model and visible in the way users filter and read results. Our eight-thousand-record review included 2,400 exact dates, 3,100 year-only descriptions and 2,500 ranges. Replacing partial values with a convenient day and month would have added unsupported precision. We instead preserved the source expression alongside the structured interpretation and its certainty. That requires product and collection colleagues to agree how a search should behave when records overlap a period rather than occur on one known day. The method does not make every date knowable. It helps the system remain useful while admitting the limits that already belong to the material it describes.

A fluent response is not sufficient evidence that a search-answer feature is ready for a collection. I ask whether the cited passage supports the claim, whether the feature declines when support is insufficient and whether the test reflects the intended use. In one 120-question exercise, eighty-eight responses were supported, twenty-two declined to answer and ten contained unsupported claims. We withheld the feature and funded further work rather than combine all non-empty responses into a success measure. Those categories describe a recorded test, not performance on every future question. Product, data and collection specialists need to agree what constitutes adequate support and how a reader can inspect it. My responsibility is to keep that evidence connected to the release decision, including the failures a persuasive demonstration might otherwise omit.