Constructive News MirrorBETA
Sign inCreate account
How it works

How Find sources works

Find sources helps you act on one of CINARA's suggestions by finding people and pages worth a closer look. It does not decide who a story should quote. It looks for the kinds of source the suggestion itself asks for, and shows what it found, where it found it, and how it searched.

The suggestion sets the kind of source

When a suggestion names who to ask, for example "an independent economist", the search looks for that kind of source and no other. When it names no one, a language model picks the one or two kinds the suggestion calls for, and if it cannot tell, you choose from the five kinds yourself. Results are grouped by kind, and the groups always appear in the same order. You can add another kind at any time.

Experts
Research or analyse the subject.
Practitioners
Work with it in practice, or have tried something similar elsewhere.
Stakeholders
Decide on it, answer for it, or represent an interest in it.
Lived experience
Live with it or are directly affected by it.
Other perspectives
Have publicly argued a different view on it.

What happens when you search

  1. Planning. A language model reads the suggestion and your article and plans two to four web searches for each kind of source, in the article's language. It may aim a search at an institution it knows works on the subject, and each kind also gets at least one search for a type of role or place, so the search does not stop at the usual names. When the subject is international, it may add up to two searches in English. It never searches for a named person, and nothing it names is shown unless a page the search finds shows it.
  2. Searching. The searches run on Google, through Serper.dev. When the search is for experts, a search in English also runs in OpenAlex, an open index of research papers, first among researchers in your country and then more widely if that finds too little. If your newsroom has switched on its archive, your own earlier coverage is searched as well.
  3. Reading. The top results of each search are opened and read, at most ten pages in all. A long page, such as a report, is read around the parts that match the search. Other results are judged from the preview the search engine shows.
  4. Checking. A language model lists the people and pages that fit one of the kinds searched and the subject of your story, copying names, roles and passages from the text. For each, it notes how closely the passage bears on the suggestion, compared only with others of the same kind, and writes one sentence on what the passage says. Code then checks every name, role and passage against the text and drops anything that is not there word for word. It also drops a sentence that adds a number or a name the passage does not contain.

Rules the code applies

  • A person is shown only when a page names them and states a professional or public role within 200 characters of the name. Someone described only by personal circumstances is never named; the page can still be listed as a place to look.
  • The authors of a research paper come from OpenAlex's own records: their name, the paper, and the institution OpenAlex lists.
  • The format, the site, the date and how a name appears (as an author, in a page title, or in the text) are read from the result by code. None of them is written by a model.
  • Within each group, what bears most closely on the suggestion comes first, and a second find from the same site comes after the other sites'. Weak finds are not removed: they come last, behind Show all. After that, people are ordered authors first, then people named in a page title, then people named in the text. For an article that is not in English, a researcher counts as an author here only if they work in the article's own country; researchers from elsewhere come last. Each research paper contributes one author: the one it names for contact, otherwise the first listed. Places are ordered with your newsroom's own archive first. Within each, the newest comes first. The format of a page never affects its position.
  • People your article already names are listed separately rather than shown as new.

What it does not do

  • It does not rank people by how good a source they would be. The sentence under a find says what the passage says, never why to contact someone.
  • It does not suggest names from a language model's memory. Every name comes from a page you can open.
  • It does not collect email addresses or phone numbers.

What stays the same

  • Always the same: the five kinds, the rules above, and the record of each search shown under "How we found these".
  • Usually the same, but not guaranteed: the searches a language model writes, the kinds it picks when the suggestion names none, and which people and pages it lists.
  • Outside our control: what search engines return on a given day, and what pages say.

Worth knowing

  • Web search tends to favour organisations that publish a lot.
  • People who live with an issue rarely appear by name. For that kind of source, the result is often places where they may be talking about it.
  • A link that opens at the passage does not work on every site. The passage shown under each result tells you what to search for on the page.

Your data

The searches go to Serper.dev and OpenAlex in the United States, and we read the pages they return ourselves. Neither receives your article. The searches are written from it, so people's names are taken out of each one before it is sent, though in rare cases one can get through. The account owner can switch the feature off for the whole newsroom, and any journalist can switch it off for themselves.

Sub-processors

How Find sources works | Constructive News Mirror