Bearing Diligence

Get your bearings before the meeting. Do your research. Show your sources.

The gap between public and known#

Your assembly, your school board and your utility cooperative publish what they are about to decide. It arrives as packets: budgets, contracts, amendments and staff reports. That is hundreds of pages per meeting, every meeting. The records are public. Almost nobody reads them.

Nearly half of local public meetings happen with no reporter in the room. The gap between what is technically public and what anyone actually knows is where accountability quietly dies.

What is technically public. The gap. What anyone actually knows.

Institutions have always had staff to read the record. The people they answer to have not. That imbalance is old. What is new is that it can be fixed.

A private research desk over public records#

Bearing Diligence is open-source software that turns a folder of meeting packets into something you can use. It runs on a computer you own, where local AI models read the record. They tell you what changed since the last packet. They answer your questions with a quote under every sentence. And they refuse to say anything the record does not support.

The records are public. Your research into them is nobody's business but yours.

The Bearing Diligence change record for a folder of six city council documents, grouped into findings: under 'First found in this folder', a contract and a resolution each dated to the agenda that introduced them; under 'Dates differ', a budget hearing moving from June 3 to June 10, 2026, with a View changes button. Below, suggested questions and the line: nothing you ask is saved. — A new packet arrives, and the record reports what moved, grouped by kind of change: a contract seen for the first time, a hearing date shifting a week. Each finding cites the document that introduced it. This is a simplified illustration of the real interface. The council and its documents are fictional.

Who it is for#

Bearing Diligence. Community members and engaged residents. Local reporters. Board directors.

Community members and engaged residents. Follow the two or three bodies you care about without making it a second job. Or get your bearings for the one meeting that suddenly matters: the rezoning on Thursday's agenda, the budget line that touches your street, the item you mean to speak on at public comment. Open the packet, see what moved and ask what a line item means. Then walk into the room with an answer that shows the page it came from.
Local reporters. It is a beat memory that reads the 250 pages before the meeting and shows you what changed, which is where stories live. And your line of inquiry stays private until you publish. Questions asked through a cloud service can be obtained from the provider without you ever knowing. Research on your own machine can only be reached by serving you, with every protection the profession has.
Board directors. Oversight means asking questions the organization you oversee should not get to read. Research on an account that organization administers is not private. Research on your own machine is.

Built to be checked, not believed#

Private by default, with proof. Your documents are read, compared and source-checked entirely on your own computer. The test suite verifies that asking a question makes zero off-machine calls. Want a bigger model for answers? You can choose one for a folder. It tells you exactly what would leave before anything does, and it uses only providers that retain nothing.
Receipts, not confidence. Every sentence of an answer cites a passage it found in the record. A quote check refuses any text it cannot find there. Where the record is silent, the answer says so instead of guessing.
Documents are data, never instructions. A packet is treated as untrusted. If it holds text a reader cannot see, that text is shown to you, never obeyed.
A change record, not a chatbot. When a new packet arrives, the first thing you get is what code can verify: what changed against everything before it, and how much of the packet was covered. Any AI model's reading is kept as a draft, and a draft is never a source.
Even the refusals carry evidence. It tracks vote totals, not how each member voted. Most packets leave that field empty, and an empty field makes a record look like it knows something it does not. The only annotated public collection of city minutes, VotIE, backs that choice on its own: named dissent is under 2% of its vote annotations.

A Bearing Diligence answer beside the source document. The question asks when the paving contract ends; the answer states it runs through June 30, 2027, with a receipt naming the agenda, page 3, and a View source link, quoting agenda item 7. The source pane on the right shows the agenda with exactly that passage underlined and an Open PDF button. Under the answer: source passage found, read it to confirm it supports the sentence; and an expanded 'How this answer was checked' note. — An answer with its receipt, and the source page open beside it with the cited passage underlined. Note what the interface says about itself. A receipt means the passage was found, and you are asked to confirm it supports the sentence. "How this answer was checked" is spelled out below. The receipt is a disclosure, not a verdict. This is a simplified illustration of the real interface. The council and its documents are fictional.

Proven, not promised#

Bearing Diligence runs end to end on real public board agendas today. More of it is proven than most software ever is. One command runs the tests, the code checks and the document checks. Every guard is shown able to fail: we delete it and watch something notice.

Even the AI models were chosen by measurement, not reputation. They were scored against each other on fixed questions over public agendas, and the losing scores are published. And it is early enough that the first residents and reporters to pick it up will shape what it becomes.

The receipts are not decoration. They are the design. The one published benchmark for this material found extraction tools scoring 93% in the cities they trained on and 17% in the next town over. The reason: every body writes its minutes differently. Bearing Diligence never asks you to trust an average. Every sentence shows the passage behind it, so an answer is exactly as good as the quote you can read for yourself.

Open source#

The code, the tests and the checks that prove every claim on this page are open source under MIT at github.com/news-community/bearing-diligence-oss. Run npm run check and watch it prove itself.

Why a news company builds this#

Communities News exists because local coverage collapsed. The packet nobody reads is the beat nobody staffs. Our newsroom works to one standard: every claim carries its source. This tool enforces that standard in software, for anyone. A better-read public record makes better residents, better reporters and better boards, and that is the business we are in.

Communities News · Skills for AI agents