Ici, l'IA n'est pas une boite noire — c'est notre phare
A manifesto for civic AI, written five days before the municipal elections of Audierne-Esquibien
The Lighthouse
A lighthouse does not choose which boats to guide. It stands on the rock, turning its beam across the water, and every vessel — large or small, familiar or foreign — sees the same light. It does not judge the destination. It illuminates the hazards so that those who navigate can make their own choices.
This is what we built Ò Capistaine to be.
Five days before the municipal elections, four lists present their visions for Audierne-Esquibien. Construire l'Avenir with Florent Lardic. Passons a l'Action ! with Didier Guillon. S'unir pour Audierne-Esquibien with Michel Van Praet. Cap sur Notre Futur with Eric Bosser. Each carries proposals on housing, schools, the port, ecology, the local economy. Each deserves to be heard on equal terms.
But citizens cannot read everything. Proposals are scattered across Facebook screenshots, PDF flyers, Instagram stories, and one participatory program co-constructed through hundreds of contributions on audierne2026.fr. The information is public but effectively invisible — buried under the noise of campaign season. No journalist has time to cross-reference every promise from every list on every topic.
The lighthouse exists to make the invisible visible.
What the Lighthouse Does
Ò Capistaine reads every published program. It remembers. When a citizen asks "Que proposent les listes pour l'economie locale ?", it retrieves what each list actually wrote, cites the sources, and presents them side by side — without commentary, without ranking, without taking sides. The answer comes with traceable references: this chunk from this document by this list.
When a citizen misspells a candidate's name — typing "van praet" instead of "Van Praet" — the system corrects it silently before searching. When a question is vague — just the word "logement" — it reformulates into something precise enough to retrieve meaningful results. These corrections happen in milliseconds, at fractions of a cent, using the smallest available model. No expensive infrastructure. No hidden cost.
Every interaction is traced. Every retrieval logged with the semantic distances that measure how well the search matched the question. Every synthesis recorded with the model that produced it. Not for surveillance — for accountability. If the system gives a poor answer, the trace shows why: the retrieval was distant, the sources were sparse, the question was ambiguous. The traces become a public audit trail.
What the Lighthouse Does Not Do
It does not tell citizens who to vote for. It does not rank programs. It does not editorialize. It does not hide information from one list to favor another.
It does not send citizen data to foreign servers. The vector search runs locally. The embedding model runs locally. When an LLM is needed for synthesis, the system tries the local model first — and only fails over to cloud providers when the local model is unavailable, with every failover logged and traceable.
It does not operate in secrecy. The code is open source, published on GitHub under Apache 2.0. The prompts that instruct the AI are visible. The documents it has ingested are enumerable. The methodology is documented. You see everything. You can verify everything.
Why This Matters
The common refrain about AI in 2026 is that it is a black box: opaque, unaccountable, impossible to trust with anything that matters. There is truth in this concern. Many AI systems are black boxes — trained on undisclosed data, optimized for undisclosed objectives, deployed without audit trails.
But opacity is a choice, not a law of nature.
We chose differently. Every decision in Ò Capistaine's architecture was shaped by a single question: can a citizen verify this? Can they see which documents were searched? Can they check whether the sources were quoted fairly? Can they confirm that no list was favored or silenced?
The answer, for every layer of the system, is yes.
This is not a technical achievement. It is a civic commitment. The lighthouse does not prove its integrity by proclaiming it — it proves it by standing in the open, visible to every passing ship, turning the same beam in every direction.
The Agents Who Learned at the School
Behind the interface, two AI agents do the work. Forseti validates citizen contributions against a charter of participation — checking for respectfulness, relevance, constructiveness — named for the Norse god of justice and reconciliation. Ò Capistaine orchestrates the retrieval and synthesis, navigating the complexity of four programs across dozens of themes.
Neither agent arrived ready. They were trained — not in the industrial sense of feeding terabytes into a neural network, but in the craft sense of refining instructions, testing against real contributions, correcting failures, and iterating until the output served the mission. They learned from the 36 citizen contributions on audierne2026.fr. They learned from the 103 documents of the participatory program. They were tested against every kind of question: vague, precise, adversarial, misspelled, follow-up.
This is Ò Capistaine's deeper identity. Not a product. A school. The agents are students who carry the provenance of their education — who trained them, on what data, with what principles, toward what purpose. When an agent graduates into production, it carries a traceable lineage that any citizen can inspect.
Three years to build this school. Not because the technology required three years, but because the trust did.
An Invitation to Every List
This manifesto is addressed to all four lists equally. Ò Capistaine serves the democratic process, not any candidate. Every list's program has been ingested with the same pipeline, indexed with the same embeddings, and retrieved with the same neutrality. The lighthouse beam does not bend.
We invite every list to examine the system. Read the code. Check the prompts. Verify that your program was treated fairly. Challenge us if it wasn't. The system is only as trustworthy as its willingness to be questioned.
And we invite every citizen to use it. Ask the questions that matter to you. Compare the programs on the topics you care about. Let the lighthouse illuminate what you need to see — and then vote with your own judgment, your own values, your own vision for Audierne-Esquibien.
The Beam Keeps Turning
On March 15, citizens will vote. The lighthouse will keep turning. Not because the election ends the need for civic transparency, but because it begins a new chapter: municipal decisions to track, council deliberations to make accessible, citizen engagement to sustain beyond campaign season.
Ici, l'IA n'est pas une boite noire. Here, it is a lighthouse — built on the rock of transparency, turning its beam for everyone, choosing no direction but illumination itself.
This manifesto was born from a conversation about whether AI can serve democracy without becoming its master. The full story of Ò Capistaine — its agents, its architecture, its philosophy — is documented at docs.locki.io. The code is at github.com/locki-io/ocapistaine. The civic platform is at audierne2026.fr.
Related: O Capistaine! My Capistaine! | The RAG Adventure Begins | Clean the Windshield