Bearing Diligence
Get your bearings before the meeting. Do your research. Show your sources.
A private civic record for an independent board director, run on local models, from Communities News.
The problem it exists for
A director's questions are her working papers: what she is checking, what she doubts, what she means to ask next. Ask them through an AI account that the organization she oversees administers, and that organization can obtain them. Independent oversight needs a place to think that the overseen cannot read.
What it is
A record that runs entirely on a computer the director owns. It holds the board's public and private material, meeting transcripts, and the local models that read them. When a new packet arrives, the record computes what changed against everything before it. When the director asks a question, every sentence of the answer carries a receipt back to located source text, checked by code rather than promised, and nothing leaves the machine.
- Local by construction. Models, documents, and questions stay on the director's own hardware. The harness tests that asking a question writes nothing that could leave.
- Receipts, not confidence. An answer's sentences cite located passages, and a quote gate refuses text it cannot find in the record.
- Documents are data, never instructions. A board packet is untrusted input: text a reader cannot see is disclosed rather than obeyed.
- A change record, not a chatbot. The primary product is what code can check: what changed since the last packet, coverage first, with any model's reading kept as a draft that is never a source.
- Even the refusals carry evidence. The record tracks vote totals rather than per-voter names, because a field that is empty in most packets makes a record look like it knows something it does not. The only annotated public corpus of municipal minutes, VotIE, independently backs that choice: named dissent is under 2% of its vote annotations, and recorded absence about 1%.
Where it stands, honestly
A working prototype, and the project's own documents lead with what that does not mean. Everything runs end to end on synthetic material; its evaluation tasks on real board packets have not been scored. The harness proves its own claims: one command runs the tests, the code gates, and the document checks, and every guard is shown able to fail by deleting it and watching something notice. Demand so far is one director's conversation, and the design puts conversations with more directors before code that assumes them.
The caution is measured, not modest: the one published benchmark for this material found extraction that scores 93% on municipalities it trained on falling to 17% on a held-out town in the same language and administrative register. Every body writes its minutes differently, which is why nothing here claims generality it has not measured on the next body's documents.
Open source
The repository is being prepared for release, including the full proposal, its decision record, and the status file that says exactly what is proven. The link will appear here the day it opens. Until then, ask us about it directly.