News Futures Knowledge Prototype 1
How can we turn knowledge into action?
This is an early test of a knowledge bank. We built it with the News Futures Working Group.
Ask it a question. It answers only from what’s inside it. But it doesn’t read its information as pages of text. It stores it as separate things joined by links. A project, the need it serves, and the place where it happened are three things. The links say how they connect. This is called a knowledge graph.
Think of a subway map. The stops are dots and the lines join them. Here the dots are needs, recipes, projects and places. To answer you, the bank travels along the lines.
For example, the Cookbook has a recipe called Care Webs: 20 or 30 neighbours who check on each other. Care Webs links to the need it serves, Well-Being. It links to the real projects it was built from, like a Tagalog phone line that young Filipino leaders set up in Dallas during COVID. And that project links to its place. So the bank can follow a thread from an idea, to the people who tried it, to where they did it.
The bank sees the world through the Civic Media Cookbook. The Cookbook starts from what communities need, then offers recipes for meeting those needs.
This is the first prototype. We arranged the information around the questions below. We think they’re good uses for a tool like this. But you can ask anything. If it doesn’t know, it should say so.
Built on Producing Civic Media (the Civic Media Cookbook) (opens in a new tab) by Jihii Jolly and Jennifer Brandel with the News Futures Working Group.
Ask the bank
What’s inside
Here is what’s in the bank. Pick a need to see its recipes. Then pick a recipe to see the projects and studies behind it.
- 10recipes
- 50projects
- 8patterns
- 54signs of a need
- 12projects with place facts
- 2,830studies, 882 linked to a recipe
1 · Needs
2 · Recipes for Agency
3 · Behind the recipe
Pick a recipe to see the projects it was built from and the studies linked to it.
8 patterns across the projects
The Cookbook also sorts its projects by what they do. Each label shows how many projects fit that pattern.
- Making Power Visible4 projects
- Amplifying Unheard Voices7 projects
- Building Civic Muscle10 projects
- Creating Shared Understanding14 projects
- Reimagining Public Resources7 projects
- Preserving Community Knowledge10 projects
- Mobilizing Collective Response6 projects
- Deepening Understanding of Place10 projects
What it doesn’t know
- Every case comes from the Cookbook's Field Guide, which describes projects 'as documented, without interpretation or evaluation'. No outcomes have been evaluated.
- No case place has fewer than 10,000 people; the smallest is an urban borough of about 14,000. Small and rural places have no close case.
- Only 3 of the 50 cases name a language; only one mentions WhatsApp.
- Population, team size, budget and urban or rural setting are not stated for any case in the book.
- Practitioner reports are not in this prototype yet.
- For research, the bank has each study's title, link and a one-line note on how it relates to a recipe, not the study's full findings.
- Nothing has been approved by a curator yet.
Review status: No item has yet been approved by a curator. Everything is as the Cookbook states it, checked word for word by machine. Bank built 2026-10-07.
From a book to an answer
How we built the bank
- Start with the CookbookThe Civic Media Cookbook. A book of community needs and recipes for meeting them.
- Take it apartAI models pulled out its pieces: 3 needs, 10 recipes, 50 projects, 54 warning signs, and more.
- Check every wordCode compared each piece with the page it came from. 1,177 checked. 5 set aside for a person.
- Join the dots257 links. Recipe to need. Project to recipe. Project to place.
- Link the research2,830 studies sorted by need. 882 linked to a recipe, each with a one-line note.
- Add the placesPublic facts from the US Census for every town in California and New Jersey, and for where projects happened.
- The bankOne file, checked and linked. This is all the room can answer from.
What happens to your question
Four stops. AI does only two of them. The other two follow fixed rules, so they do the same thing every time.
Stop 1 · AI
Pick where to start
An AI model reads your question and a list of everything in the bank. It picks the stops to start from: which needs, recipes and projects fit.
Stop 2 · rules
Walk the lines
From those stops, the bank follows its links. Recipe to need. Recipe to projects. Project to place. If you picked a place, it compares sizes. It gathers facts and gives each one a label. If something is missing, that becomes a fact too.
Stop 3 · AI
Write the answer
A second AI call writes the answer. It may use only the facts from stop 2. It must point to the fact behind each claim.
Stop 4 · rules
Check it
Rules check every label points to a real fact. They remove long passages copied from the Cookbook. They keep the answer under 360 words. A claim with nothing behind it is cut.
Your answer · every part carries one of three labels
- Grounded in the bank. The bank says this. The label beside it shows where.
- Informed by the bank. The bank’s own reasoning from what’s in it. The bank doesn’t say it outright.
- Beyond the bank. The bank has nothing on this, so you may want to look elsewhere.
Open-weights models with zero data retention
The best-known AI models are called frontier models. ChatGPT and Claude are two. They are the newest and most powerful. Their insides are secret. You can only rent them from the company that made them.
We use open-weights models instead. A model’s weights are the billions of numbers that store what it learned. Here they are published. Anyone can download them and run them. So we aren’t tied to one company, and someone else could run this bank after us.
Our jobs are narrow. Read a book carefully. Pick from a list. Write from facts we hand over. Open models do these well, and cheaply. Turning the whole Cookbook into data cost about 4 cents.
They keep nothing. We only use companies that promise to keep no copy of what we send. Your question is used to write the answer, then thrown away. This promise is called zero data retention.
Answers your questions · built most of the bank
gpt-oss-120b
Made by OpenAI and released with open weights in 2025. “120b” means about 120 billion weights. Run for us by Groq, a company with very fast chips.
Helped build the bank
Qwen3 235B
Made by Alibaba’s Qwen team. It read 9 worksheet pages the first model found hard. Run for us by Nebius, in Europe.