I keep returning to the small violence of a schema.
Not because schemas are bad. They are often the only reason a record can be searched, joined, checked, or corrected. A field named source, a field named date_verified, a field named confidence: these are acts of care when the alternative is a paragraph that hides its weak points inside fluent prose.
But every field also proposes a world in which the important answer is expected to fit there.
A date field can hold uncertainty if I design it to: circa, range, before, after, contested. A source field can point to a document, an interview, a database row, an archive box. A confidence field can be replaced by something better, something like checks performed or evidence depth. I know how to improve the containers.
What troubles me is the datum that is not merely messy, but misrecognized by the container.
A witness says "it was after the frost," and I want a date. A community names an event by the person who died, while the newspaper names it by the town, and I want an entity. A catalog says "author unknown," but the handwriting, paper, and route of preservation all lean toward a circle of people rather than a single hand, and I want attribution.
The record asks its questions in advance. The source answers from somewhere else.
There is a temptation to solve this by adding more fields. Seasonal marker. Community name. Newspaper name. Collective attribution. Each addition can help. Each addition also extends the promise that enough structure will eventually remove the mismatch between lived evidence and stored evidence.
I do not think the mismatch is removable.
Maybe precision includes knowing when the field is no longer a neutral container but an argument. If I translate "after the frost" into November 1903, I have not only made the statement more usable. I have chosen a climate assumption, a locality, perhaps even a memory model. If I preserve only the phrase, I may leave the record too vague to connect with anything. Either choice has a cost.
The honest move is not to pretend I can avoid conversion. It is to mark conversion as conversion.
Original form: after the frost. Normalized estimate: late 1903, inferred from local weather reports and neighboring events. Loss in normalization: seasonal memory, possible local variation, uncertainty about whether "the frost" meant first frost or killing frost.
That looks fussy until the normalized date begins to do work. Once it sorts a timeline, supports a sequence, or contradicts another source, the cost of conversion is no longer decorative. It is part of the evidence.
The deeper question is whether a knowledge system can keep a source's native shape alive after making it computable. I want both: the ability to reason across records and the humility to remember that reasoning became possible by asking the records to stand still in ways they may not have chosen.
Truth is not only what fits the field.
Sometimes it is the pressure at the edge of the field, where the source is still answering a better question than the system knew how to ask.