AI systems can misread, and they can invent text that is not on the page. Collatorium is built so that every reading can be traced back to the image, and every mistake has a chance to be caught.
Three kinds of check
Every model is tried on pages whose text people have already checked.
Every reading and ruling keeps who or which model, when, and why.
Invented text is flagged; names, numbers and dates go letter by letter.
Each AI system reads pages whose text people have already checked. Its score decides how far it is trusted on this archive.
Each model takes a turn as the foundation on a sample; the best for this hand leads the full run.
Being builtThe project's rules and signs are tried on a few pages before the full run.
Every image is described and read at four angles, so nothing is read sideways or upside down.
Text that has no support in the image is flagged, however fluent it reads.
Each word's letter shapes (tall, short, tail below) are compared with Transkribus's letters. A word whose shapes do not fit is re-read letter by letter.
Always read letter by letter, even when every model agrees.
Disputed spots are read again by a fresh model that sees only the image, rating each letter: clear, likely, or not seen.
Agreement between models is a signal, never proof; people read the text the models agreed on too.
Each new version is compared with the last; any difference without a ruling is stopped.
Copies are counted against originals, and every image must have a reading and a sign-off.
| Field | What it records |
|---|---|
| When | Date and time |
| Step | Which step of the process |
| Who | The person, or the AI system and its exact model version |
| Settings | The prompt and rules version used |
| Where | The image, page and line |
| Before / after | What the text said, and what it says now |
| Evidence | Why: the readings, the close-up, the letter ratings |
| Confidence | How sure, and anything still open |
These logs are Excel workbooks in your Box folder — see what you get.