Market Evidence Desk: Research With Receipts

Paul Ayoade — Tue, 29 Sep 2026 11:53:47 +0000

This is a submission for the Sanity Challenge, Path One: Ship an Agent That Queries Real Content.

What I Built

Market Evidence Desk is a public research desk for crypto investor protection questions. It links original publications to dated claims about proof of reserves, custody, deposit insurance, stablecoin redemption, and the gap between international recommendations and local implementation. An exchange may say customers can check inclusion in a reserve snapshot, while accounting and regulatory sources explain why that snapshot cannot establish full liabilities or solvency. SEC staff and a Commissioner also take different positions on what reserve reports demonstrate. Flattening those positions into a single confident answer would lose the point.

The site lets visitors inspect published source records, follow each claim to its research question and original URL, compare dated views, and ask a scoped AI agent. As of September 30, 2026, the published Sanity graph contains 101 research questions, 112 claims, and 22 source records representing 21 distinct source URLs. The evidence atlas shows the records in manageable batches. Source scope notes explain whether a document is investor education, analysis, or a recommendation to regulators.

The agent returns a research draft with source links and the names of the Sanity Context tools it used. It declines live Bitcoin prices and buy/sell requests because its sources do not include a current market feed. Generated drafts remain subject to human review.

Demo

Live desk: market-evidence-desk.vercel.app
Evidence atlas: Browse research questions and their linked evidence
Compare evidence: See attributed positions side by side
Sourcebook: Inspect original sources and claims
AI agent: Try the disagreement case
Video walkthrough: Watch on YouTube

Instructions for Judges

The public app does not require sign-in:

Start on the home page, select Show the disagreement, and press Ask the sources.
Compare the agent’s cited answer with the original sources and the guided comparison.
On the Research desk, switch to the question about worldwide protections. The desk links 2023 international recommendations and 2025 implementation reviews as different kinds of evidence.
Try the live-price boundary case to check how the agent handles a question its sources cannot answer.

The agent can take up to two minutes. The guided comparison and published source graph can also be explored independently. The Fictional demo on the Research desk is explicitly separate from published records.

Code

GitHub repository · Architecture and setup · Scope and limitations

How I Used Sanity

The production dataset models source, marketEvent (a research question), and evidenceClaim as separate linked documents. Each claim points to a source and a question. It also records a date, review state, and stance such as supports, context, or conflicts. The public pages query published records and show each position with its origin and limits. That structure helps readers distinguish a company’s description of its own process from independent assurance.

Sanity Context points at both the dataset and the Market Evidence Desk Sourcebook Knowledge Base. The Sourcebook selection is capped at 140 source and claim documents, with a separate SEC investor-alert page. Its September 27 rebuild succeeded with 13 cited entries and zero open issues. The proof-of-reserves entry preserves a source-scoped correction that distinguishes Kraken’s description from PCAOB and SEC cautions. The import total, Knowledge Base build count, and current public graph count measure different things; the full published graph remains available through the public pages and GROQ.

The Python agent calls groq_query to retrieve the linked published graph and knowledge_base_read to retrieve relevant Sourcebook entries before answering. The server withholds an answer if either required retrieval is skipped. Visitors can ask a scoped question on the public site, while credentials remain in server environment variables. The guided comparisons read Sanity records directly and are not AI-generated.

One hard case concerns two SEC publications from April 4, 2025. The Division of Corporation Finance staff described an issuer practice involving stablecoin reserve reports. Commissioner Caroline A. Crenshaw questioned whether such reports establish adequate backing. The desk attributes both positions without treating either as a Commission rule or a finding about present-day solvency. It also distinguishes FSB and IOSCO recommendations from rules adopted by individual countries.

All 112 published claims currently carry source-checked status with an excerpt, locator, reviewer name, and review date. Those fields record an editor’s source-check attestation. They do not constitute an independent audit of every passage or an endorsement of an AI-generated answer.

Sanity Project Details

Project ID: cxjysvlq
Dataset: production
Public dataset: Query published source, question, and claim records
Sanity Studio: phoniex-market-evidence-desk.sanity.studio
Sourcebook Knowledge Base ID: kbu9WNgZ9ocF

Agent Session

The public ten-question agent session preserves the questions, full answers, reported Sanity Context tool calls, and grading notes without private headers or keys. Eight questions passed and two failed the initial check.

After fixes to retries and source validation, I reran the same ten questions. All ten returned HTTP 200 on the first call and passed the original core criteria. This small manual evaluation is not a reliability guarantee. These are public repository transcripts rather than a DEV Agent Session embed.

What I Learned

Structured content matters most when sources disagree or address different scopes. Dates, authors, stances, and source references let the interface show what each publication can establish. Sanity Context lets the agent read both linked records and the Sourcebook, while the review state reminds readers that a generated explanation is a draft. A useful answer also has a visible boundary: historical evidence cannot supply a current price or an investment recommendation.

DEV Community: Paul Ayoade

MARKET EVIDENCE PROJECT