understories.io • translating climate doom into agency

—————————————

1 . Data and updates

I have not shared data since August. The overall trend remains consistent with the project’s early months, but new usage patterns are emerging alongside the content being published:

• Over the past two months, visits have come from more than 60 countries. Alongside the US, UK, and France, countries showing growing traffic include Canada, Switzerland, India, Singapore, China, and Indonesia.
  
  

• Last month’s most-visited pages were the homepage (25.4%), the Library (20.9%), and Test Your Message (15.3%). After a pause, the latter doubled in one month after the tool went live and became automatic.
  
  

• Test Your Message is live. If you have not yet tried it, please do. The beta version allows you now to test up to 15 messages per month. The system is GDPR-compliant and uses double opt-in. Evaluations are sent from a dedicated address, warren@understories.io. It is not a robot, but a separate email dedicated to sending evaluations.

2. The tech and its surprises 

The system and codebase are fully operational. Unsurprisingly, moving from prototype to the new orchestration brought surprises and choices. For anyone interested in exploring or building a similar system, here are a few points I am happy to share:

• Evaluation: Numeric scores from the prototype have been replaced with five colour bands. Numeric scores still exist internally, but are no longer delivered to users. The evaluation of the final revised text has also been removed, allowing users to decide how to resolve tensions within their own text. The system suggests; you decide.

• Vector database: The prototype did not embed the full library of 50+ reports because of weight and cost constraints. The live system now does, not as original PDFs, but as structured collections with guidance. It is hosted on Qdrant, a German open-source vector database.

• Retrieval design: Understories uses retrieval-augmented generation. The beta retains the prototype’s three-agent orchestration and adds a synthesiser node that helps ensure that the final evaluation remains coherent, balanced, and centred on the user’s original goal and call to action.

• Tools and models: The pipeline is built with Haystack, a German open-source Python framework, while pre- and post-safety tests used Giskard, a French open-source tool. Understories’ live system currently uses Claude Haiku 5.5; further LLM testing is ongoing.

3. A perspective from outside 

It was a pleasure to speak with Catherine Mackie from the Thomson Foundation about the Climate Crisis Toolkit for Media in Tanzania, which is part of Understories Library. What I appreciate most about this resource is its focus on building capacity.

• It is practical, not just informative. The toolkit combines storytelling guidance with Tanzanian case studies, interactive exercises, safety guidance, and linked resources. Thonsom Foundation designed it in partnership with the Tanzanian Media Foundation, Climate Action Network Tanzania, among others.

• Its local grounding makes it transferable. Although designed for Tanzanian media, it taught me, as a French communicator, a great deal about the realities faced by East African journalists, including what it means to report in conflict-affected contexts. Its lessons on advocacy, policy, legal issues, and audience trust are relevant well beyond Tanzania.

• It aligns with Understories’ ethical approach. By helping journalists avoid false balance, jargon, and disempowering narratives—and by treating “missing voices” (gender, ordinary people versus officials, and marginalised groups) as an equity issue in storytelling—the toolkit supports accurate, accountable, and more inclusive climate storytelling.

  
  The toolkit is available free on Thomson’s Journalism Now e-learning platform in English, with a Swahili version accessible through a translation button.

See the article here Thomson toolkit added to virtual hub built to strengthen climate storytelling”.

—————————————

On ethics and environment

Mongabay has published a commentary, “AI-centered conservation efforts can only be ethical if Indigenous people help lead them,” by McKalee Steen and Magali de Bruyn. The authors argue that AI in conservation can only be ethical when Indigenous peoples help lead it—not simply consent to tools designed elsewhere.

They call for Indigenous digital sovereignty: communities governing the collection, ownership, and use of their own data. Their examples include an advanced sea-ice modelling and forecasting service led by Alaska Natives.

—————————————

4 . What’s next

• Explorations section: During the revision of the methodology paper, reviewers suggested that it could better address the three audiences it is intended for. I am therefore developing it into three papers: a ReadMe for AI builders (already published in the private Github repository), a methodology paper for researchers (being rewritten), and an essay for practitioners. The essay will be sent ahead of the next Reading List issue on Thursday 22 October.

• Making the project sustainable: Understories has been accepted into MakeSense, a Paris-region-based six-week fundraising programme. Among its many activities, MakeSense runs an incubator track helping associations and start-ups diversify their funding. I am pleased by the acceptance because of the institutional signal it sends: a go-ahead for the next phase. The aim is to keep evaluation free for under-resourced climate advocates, so over the coming weeks I’ll be working on the funding strategy that makes that possible. 

—

You can subscribe here: Join the Monthly Updates