Sources enter Sanity
Each record keeps a title, URL, publisher, publication date, and source type. That is the receipt layer.
05 / HOW IT WORKS
A small evidence system with a clear chain of custody.
The project connects published sources, research questions, evidence claims, human review, and an AI agent that must show its work.
Claims attach to questions
A claim says what a source supports, conflicts with, or provides as context for a research question.
Uncertainty stays visible
Questions carry a review status. The system does not pretend that model synthesis is an approved conclusion.
The agent reads before writing
The Sanity Context agent queries the dataset and Sourcebook, then returns an evidence-grounded draft with tool receipts.
THE CORE RULE
Every strong sentence needs a traceable receipt.
If the dataset cannot answer a question, the correct output is a clear boundary, not a confident guess.
SOURCEWhat was published?
Read the original document.
CLAIMWhat does it actually say?
Keep the stance and date visible.
REVIEWWho checked it?
Drafts wait for a human checkpoint.
A PRACTICAL READING CHECKLIST
What would make a brief dependable?
1. Open the original
Check title, publication date, publisher and whether the document is a staff view, a regulator recommendation or a company account.
2. Trace the claim
Follow the source to its linked research question. Look for supporting, contextual and conflicting statements before repeating a conclusion.
3. Mark what is unknown
Identify the jurisdiction, review date and unanswered parts. The absence of a live feed or issuer audit means current safety and price questions remain open.
WHERE THE DATA LIVES
Two Sanity layers, one review boundary.
Published source, question and claim documents form the structured graph shown across the public pages. The Sourcebook Knowledge Base organizes source material for retrieval by the AI agent. The two are related, but a Knowledge Base entry does not by itself approve a research conclusion.
What the agent must do
The server asks Sanity Context for both a GROQ query and a Knowledge Base read. It then returns the generated draft with a tool receipt and links only to sources in the published graph. Reviewers still need to inspect original passages and any Sourcebook issues.
Try a scoped question ↗Contest position: the novelty is not “an AI that talks.” It is an AI research workflow that exposes provenance, disagreement, and the limits of its evidence.