Every probability says where it came from.
Each cell of each conditional table records its origin, the model and template version that produced it, and how many samples stood behind it. A number you cannot source is a number you cannot review.
Most models record provenance on the model. That is enough to say a model was elicited and not enough to review one, because the interesting question is per-number: which of these came from data, and which from a language model's reading of a paragraph.
Four origins
Stored as a tag on the row, not as a note on the model.
- Elicited
llm_elicited A language model produced the number while turning a description, document or table into structure and probabilities.
How to read it: A starting hypothesis. It is a claim about the domain made by something that read the material — useful, inspectable, and not evidence.
- Expert
expert A named person replaced or supplied the row, with their rationale recorded beside it and a new revision created.
How to read it: The strongest form of judgment available when no data exists, and the easiest to misremember later — which is why the rationale is stored rather than recalled.
- Fitted
fitted The row was calibrated from observed counts, under a documented procedure, with the sample count recorded.
How to read it: Backed by data. It records how many observations stand behind it, so a row fitted on nine rows is distinguishable from one fitted on ninety thousand.
- Mixed
mixed Some levels came from a fit and others from elicitation, which happens whenever the data covers part of a variable's range.
How to read it: The tag that matters most in practice, because it is the one a summary would round away. It is the reason provenance is per row.
Provenance describes where a number came from. It is not a judgment about whether the number is right, and a fitted row does not by itself establish a causal relationship.
What it lets you do
Three things follow from per-row provenance that do not follow from a note on the model.
Act only on what is calibrated
A projection returns the largest sub-network in which every row is backed by a fit, closed under its parents. Ask a question inside that region and you are answering from data; ask outside it and you know you are not.
What this does not establish: It narrows the question rather than improving the answer. A smaller calibrated region can be too small to contain the decision you actually have.
See where elicitation disagreed
Where the ensemble produced materially different structures, the edge is reported as contested rather than resolved by majority. A disagreement between ways of reading the same document is a signal about the document.
What this does not establish: Contested is a review flag, not a calibrated probability that the edge is wrong. It says several readings were defensible, and stops there.
Replace a number and keep the reason
An override replaces one probability cell with a named judgment and its rationale, and creates a revision rather than mutating the model. The previous value remains readable, along with who changed it and when.
What this does not establish: An override is recorded, not validated. The model will use whatever number it is given, and the record shows the change without judging it.
Why per-row matters for the decision
A model assembled from a report, an expert and a claims extract has three different kinds of support inside it. If that distinction only exists on the model, then every answer computed from it carries the average of those supports — and the average is the one thing none of them actually is.
With the tag on the row, the distinction survives into the answer. A stored analysis records which rows it read, so a reviewer can see whether the number they are being shown rests on the part of the model that came from data.
This is also what makes a later revision legible. When a fit replaces an elicited row, the change is visible as a change in support rather than only as a change in a value.
If you have a decision to make and a number someone will hold you to, the question is not whether we can produce a number. It is whether you can say where it came from.