For the person who has to answer a question about the whole library, not about one file.
Here, an impossible question is one that is impractical to answer by reading the documents one after another, at the scale and in the time available. It is the name of a way of asking, not a promise.
Questions that appear when documents are read together
Answered with an address you can check
Some questions are not about any single file. Which of forty proposals still carry a hidden slide? Between two versions of a deck, which numbers in the text changed? Which contracts in a folder still hold comments, and what do they say? SazAI Corpus reads each document once, element by element, and keeps the record: what each file holds, which file SazAI Corpus produced from which, and the comparisons already run. The assistant you use asks over that record, and each finding can come with the document and the element it rests on, so you can open it and check.
What makes a question impossible
It needs at least one of three things
All of them, with the count. “Of all the proposals, which…”, “none of them”, “which has the most”. A search that returns the closest passages cannot say “all”. A query can be paged through to the end of the scope you set, and the answer says, beside it, what was not read. Example: of the forty proposals in this folder, which carry a hidden slide?
The library’s memory. The versions SazAI Corpus produced from a file, and the comparisons it ran between two files. These records come from Corpus itself, not from the files. Example: which versions were produced from the same source file, and which pairs have already been compared?
The layer under the text. Formulas, hidden slides, comments, image descriptions, signature fields. A tool that keeps the text alone drops this layer before anyone asks. Example: which contracts still hold comments, and what do they say? Signature fields are read as they are declared: SazAI Corpus does not verify a signature.
A question that needs none of the three is an ordinary question, and it gets an ordinary answer.
How to read a finding
What a finding in this mode should carry
- The bridge. Two documents, or two versions of one, and the element that links them: a formula, a version, a sheet that lines up, a shared term. No bridge, no finding.
- Two explanations at least. Such as a shared source, a common template, or chance, and the item that tells them apart. Evidence that fits every explanation decides nothing.
- The coverage. How many documents were covered, which were left out, and what was not read.
- The address. The document, and the slide, page, paragraph or cell.
- The next check. The one whose opposite result would most change the conclusion.
An inconclusive result is a result. Ask for these points in your question; a finding that lacks one of them is a lead, not a finding.
What SazAI Corpus does here
Scenes scored in a run
No question in this mode has been scored in a run yet. Scenes appear here when one is.
What SazAI Corpus does not do
In this mode, today
Corpus does not:
- promise that something will be found;
- say whether there was fraud, intent, or a breach of a rule;
- read a missing link as proof that two documents are unrelated, or two similar texts as proof that one was copied from the other;
- read an order in time as a cause.
Corpus reads and records; your AI reasons over it; you decide. The fields where this applies: Audit, Compliance, Financial services, Legal, Life sciences and Real estate. To talk to us: contact.
Reading list
The basis of the method
The method is built on this published work. Citing it does not mean its authors endorse SazAI.
- Swanson, D. R. (1986). Undiscovered Public Knowledge. The Library Quarterly 56(2), 103 to 118.
- Pirolli, P., Card, S. (2005). The Sensemaking Process and Leverage Points for Analyst Technology. International Conference on Intelligence Analysis.
- Heuer, R. J. (1999). Psychology of Intelligence Analysis, chapter 8, Analysis of Competing Hypotheses.
- Buneman, P., Khanna, S., Tan, W. C. (2001). Why and Where: A Characterization of Data Provenance. ICDT 2001.
- Howard, R. A. (1966). Information Value Theory. IEEE Transactions on Systems Science and Cybernetics 2(1), 22 to 26.
- Kastrin, A., Cestnik, B., Lavrač, N. (2025), and Alkan, A. K. et al. (2025): surveys of literature based discovery and of hypothesis generation.