Glossary — Strategic Agent

Read this page if a word on another page stopped you. Every term below is defined in plain words, in the sense this product uses it, and nothing else. Where the word names something that has been measured and found short, that is said here too, because a definition that hides a defect is not plain.

The records

six words about what goes in

Everything starts with paper. Before the product can answer anything, it has to read what you have and keep an account of how the reading went. These six words are the vocabulary of that account. In the proving run they described 56,687 files, of which 53,776 came out readable, 98% clean, with two held back for a person.

Term	Meaning	Where it matters
corpus	The whole set of your records the product has read: the originals, the readable copies made from them, and the index built over those copies. One set belongs to one owner and is never merged with anyone else's.	Your own records
chunk	A short run of consecutive lines cut from one readable document, small enough to be scored against a question on its own. The index holds chunks, and every chunk records which document it was cut from.	How it answers
manifest	The ledger a conversion run keeps: one entry per unique file, with its path and current status. It makes a run resumable, and it is what lets the product say, file by file, what was read and what was not.	The manifest
quarantine	The status given to a file that could not be converted after a bounded number of attempts. A file in this state is set aside for a person to look at, never silently skipped, and its original is left untouched.	Never loses a file
OCR	Optical character recognition: software that reads printed text from a picture of a page. Scans and photographs need it before they can be searched. The product records a confidence for every page and runs it only on your machine.	What it cannot read
provenance	The recorded history of a readable copy: the original's path, the method used to read it, the status it ended in and, for scanned pages, how confident the machine reading was. Where a document came from, not who wrote it.	Provenance

Two of these carry a number worth knowing. The archive OCR run to date is 7,996 documents and 35,633 pages, every page read on the machine with nothing sent out — an actual. And the manifest counts unique files, never rows, because a retry loop once made one look 3.6 times larger than the files it described.

The answer

four words about what comes out

An answer here is not a sentence about your documents. It is a line out of one of them, with the document named and the place in it given, or it is a refusal. The cap is part of the promise: at most three lines and at most two pieces of proof, so that everything returned can be checked.

Term	Meaning	Where it matters
grounding	The rule that every answer is a line present in your documents. Nothing is composed: text is found, scored and returned as it was written. A question that cannot be answered this way is declined, not answered from anywhere else.	Grounded or silent
citation	The proof attached to an answer: the document's identifier, the line numbers, and the text exactly as it appears there. At most two accompany any answer. Open the named file at those lines and the same words should be there.	The citation
locator	The part of an answer's proof that says where in the document the quoted lines sit: a line number, or a span of them. With the document's identifier it is enough to open the file and find the text yourself.	The citation
abstention	The product's refusal to answer, given as one fixed sentence when your records do not hold enough to answer with confidence. It is the designed outcome for a question the records cannot support, not an error and not an apology.	When it abstains

"Your records don't contain enough to answer that with confidence."

That is the whole of the abstention, word for word. It does not vary with the question and it does not soften into a partial guess.

The people

two words about who

The last two words are about people rather than files: whose records are whose, and what a person must do before anything the software proposes becomes part of the record.

Term	Meaning	Where it matters
tenant	One owner of one set of records. Every read and write in the store must name the owner it belongs to, and the software refuses if it does not. Two owners on one machine cannot see each other's records.	Security
ratify	To accept, as a person, something the software has proposed, so that it enters the record. Until a proposal is accepted it stays a proposal, whatever it says. The software never does this step for itself.	The human gate

A definition is not a measurement

Abstention is defined above as the designed outcome. Measured once, across three of the proving corpora, it failed to happen 45 times when it should have. A locator can currently report one continuous range of lines for text stitched together from lines that are not adjacent — a known defect in the citation. And OCR is defined here as the step scans need, not the step they have all had: in one of the proving corpora, none of its 883 scanned candidates has been read yet. Each of these is on the record page, with what would close it.

Read the record ▸

three doors

What it is

The plainest statement of the product: the four things it is built on and the two halves it is made of.

What it is ▸ What it is