Step 1: Intake
Intake accepts four kinds of input. For each it records what code measured, what the model inferred and what is unknown.
| Input | How it is read | Limit |
|---|---|---|
| Link to an article or report | The model fetches the page and saves the text with the URL, publisher and date. | The fetch tool can summarise long pages, so quotes cannot be checked against the page. Intake flags this. Above exploratory stakes, ask for the document as a file. |
| Content pasted from an article or report | Saved as text. | Publisher, author and date are unknown unless the user says. |
| Typed text | Under 40 words it is the request. Longer text is material. | A request with no material produces a question asking for material. |
| Collection of documents, such as transcripts | Files or a folder (txt, md, html, docx, pdf, csv, tsv, xlsx). | Unreadable files are listed and the user is asked to paste the text. |
The output is the brief. Each part is marked stated, inferred or missing.
The request. One request, which can have several parts or be open-ended, such as "tell me what insights I can get from this article". It is treated as one request and gets one set of techniques.
The settings.
| Setting | Question to the user | Values | What it changes |
|---|---|---|---|
| Stakes | What will the result be used for? | Exploratory: the result decides what to look at next. Decision-informing: one input to a decision that can be revisited at moderate cost. High-stakes: a costly or hard to reverse decision, one that affects rights, safety or finances, or an authoritative publication. | Higher stakes add library checks. |
| Strength of statement needed | What is the strongest statement you need to make? | Describe. Show a link between things. Predict, without claiming a cause. Estimate the effect of an action. Explain how or why. | It adds library checks, guides the choice in step 3 and is reported in the answer. It never removes a technique. |
| Kind of finding wanted | What kind of finding are you after? | Pattern, difference, contradiction, cause, trend, gap, risk, option ranking. More than one can apply. | It guides the choice in step 3. It never removes a technique. |
For an open-ended request the kind of finding is left empty and the strength of statement defaults to describe.
The description of the material. One value for each property that a requirement tests, recorded per source where the property belongs to a source. Code fills content type, structure, level of detail available, unit of analysis, number of sources, whether sources can be joined, amount of material, document type, language, signs of personal and sensitive information, and known data problems (duplicates, share of blanks). Units of analysis are offered broadly, because the unit depends on the request: text is offered as document, claim and topic, and as source when there are several.
Questions to the user:
- At most four at once, in this order: missing material or request, stakes, strength of statement, personal information, kind of finding, joining of sources, duplicates. The rest are not asked in that round.
- Each question has a recommended answer where one exists and a one-line reason.
- Time structure, how the sample was chosen, origin, the producer's interest and whether sources depend on each other are never asked. They stay unknown unless the material states them.
- Security classification and licence are not asked, because no technique uses them.
Inferred without asking, and shown when the user confirms the brief: no personal information, for report-type documents with no emails, phone numbers or sensitive words; strength of statement and kind of finding, from the wording of the request.