Hello,

Today marks five months since Understories' research pilot went live. It also marks the release of the public version, with a new story: Marcel's.
Keep in touch,

Caroline

—————————————

1 . Discover Marcel’s story and why this message scored a C

Marcel is a French farmer: see in LinkedIn the draft message, the evaluation, and the revision. Then try it on your own message.

2. What shipped, from pilot to public version

How it works

The system uses retrieval-augmented generation (RAG) to connect a language model with the research library. In simplified form, the process is:

1. A user submits a draft message with its audience and desired goal.

2. The system evaluates the message across the eight narrative criteria, retrieving material from the library, and scoring against that evidence. 

3. The system produces criterion-level observations, citations, a revised message, and unresolved tensions.

4. The user decides whether and how to revise the message.

3 . What changed

In comparison with the research pilot that worked “manually” from May to mid-July, the changes made include architectural, technical, and design changes.

• Test Your Message, upgrade 

Users paste their message and must confirm their email addresses to receive their evaluation. The message passes through a pipeline, with upgrades:

• a node synthesizer, “Warren”, coordinates the agentic work; 

• the draft message is scored against the full 50+ research library studies;

• The A-E bands replace the 1-10 grading with A=10-9, B=8-7, C=6-5, etc.

• citations focus on the lowest criteria and on explaining what changed and why.

• the refined message is not scored; open tensions are left for the user to decide.

The current beta is capped to one evaluation per hour and four evaluations per month for now. Tests are still ongoing. 

• The Privacy Policy: It has been updated here and applies now to the Evaluations, the Monthly Updates, and the Reading List.  The service and the database that stores submissions run on an OVHcloud server in Strasbourg, France. Understories’ policies are aligned with GDPR, French CNIL, and EU AI Act standard practices. 

• The FAQ: It has been updated to match the public version and to mention relevant information about the upgraded system. The Ethical framework and the Environmental Impact are discussed independently, at a slower pace, and with external parties.

—————————————

4 . What’s next

• Methodology Paper: The methodology presents the system architecture.  is now being reviewed by researchers, practitioners, and AI Ethics lawyers. Again, a big thank you to the reviewers who bring expertise and ask relevant questions.

• Users’ feedback and testimonials: Now that the public system is open, I’m collecting feedback. What works well, what works less well. Please share feedback and share testimonials too!

• The Reading List’s new season [after a break]: Thursday's edition #8 is the last issue of this season, focused on the Research Library. A new series starts at the end of October. Sign up if you’re interested.
  
  

Climate is not only a policy problem. It's a story problem. 

The research is there; the ethics are there. Understories is an attempt to close the gap between research and practice and tell the climate story better. 

Thank you for reading until here. 

• ——

The monthly update is shared by email first to users, researchers, AI and comms practitioners I happen to meet while building Understories.

You can subscribe here: Join the Monthly Updates