A crypto research agent that can’t show where a claim came from is a fast way to get confidently wrong. For DeFi, a tidy answer is not enough: you need to trace important claims back to current, relevant sources.
Treat the agent as a research assistant, not an oracle. Ask it to gather evidence, then check the evidence yourself before using the result in a trade, risk review, or governance decision.
What a useful source check looks like
A citation is useful only if it gets you to evidence that supports the claim. A link to a project homepage doesn’t prove a current fee, and a link to a governance forum doesn’t prove a proposal passed.
For each important claim, check four things:
- Origin: Does the link open the original source, such as a protocol document, governance proposal, contract, or data dashboard?
- Support: Does that page actually say what the agent claims?
- Freshness: Is it current enough for the decision? A contract address can remain valid for years; a lending rate can change in minutes.
- Scope: Is the evidence about the right chain, market, version, and time period?
Ask the agent to separate facts from interpretation. “The proposal passed” is a checkable fact. “This makes the token undervalued” is an interpretation that needs more assumptions and evidence.
A practical workflow
Start with a narrow question. “Research this protocol” invites a generic summary; “Did this proposal change the liquidation threshold on Arbitrum?” gives the agent something you can verify.
Then require evidence in a consistent format. For every material claim, ask for the exact source URL, the relevant passage or data point, its date, and what the source does not establish. That last part matters: a governance post may describe a proposed change, not a change that has already been executed onchain.
Open the sources yourself. Prefer primary evidence in roughly this order:
- Onchain transactions, verified contract code, and official chain explorers.
- Official documentation and protocol dashboards.
- Governance proposals and execution records.
- Independent analytics and reputable research.
- Social posts and third-party summaries as leads, not proof.
No source type is automatically correct. Documentation can lag behind a deployment; dashboards can have indexing delays; an explorer can show a transaction without explaining its practical effect. Cross-check consequential claims against a second source with a different failure mode.
Worked example: checking a lending-market change
Suppose you’re deciding whether to deposit into a lending market on Arbitrum. An agent says: “The protocol lowered the collateral factor for Asset X from 80% to 70% yesterday, increasing liquidation risk.”
Don’t paste that sentence into a group chat as fact. Give the agent a verification task:
Check whether Asset X’s collateral factor on Protocol Y, Arbitrum, changed from 80% to 70% on 2026-10-17. Return a table with: claim, source URL, exact evidence, timestamp, and confidence. Prioritize the official governance proposal, execution transaction, and current protocol dashboard. If you can’t verify a detail, label it unverified. Keep facts separate from your interpretation.
Then check the answer step by step:
- Confirm the proposal. Open the linked governance page. Check that it names Asset X, the Arbitrum market, and a change from 80% to 70%. A discussion or draft proposal only shows intent.
- Check execution. Find the transaction or execution record. Confirm it happened on the relevant chain and after the proposal passed. If the proposal passed but the execution transaction is missing, the change may not be active.
- Check the live parameter. Inspect the official market dashboard or contract state. Verify the current value is 70%, and confirm the dashboard is showing Arbitrum rather than another deployment.
- Check the date. A page’s publication date, a transaction timestamp, and the time you checked current state are different things. Record the relevant one for each claim.
- Separate fact from impact. The parameter change may reduce borrowing capacity for new positions, but the effect on your deposit depends on your position, other risk settings, and the protocol’s liquidation rules.
Your notes might look like this:
| Claim | Evidence to record | Status | |---|---|---| | Proposal requested 80% → 70% | Proposal URL and parameter passage | Verified if the text matches | | Change was executed yesterday | Transaction URL, chain, timestamp | Verified only if execution is confirmed | | Current factor is 70% | Current state or dashboard, checked time | Recheck before acting | | Liquidation risk increased | Your position plus protocol risk rules | Analysis, not a source fact |
If the agent cites only the proposal, its answer is incomplete. If it cites a dashboard but no chain or timestamp, it may be looking at the wrong market or stale data. Ask it to fill the gap; don’t let a confident summary substitute for missing evidence.
What to use, and when
A general-purpose agent can be useful for finding relevant documents, comparing statements, and turning research into a checklist. It is less dependable for live protocol state unless it has access to current, chain-specific data.
| Approach | Pick it when | Tradeoff | |---|---|---| | Research agent with citations | You need a quick map of documents and claims | Links can be weak, stale, or misread | | Block explorer and official dashboard | You need to verify a transaction or live parameter | You must know what to inspect | | Manual search | The question is narrow and primary sources are easy to find | Slower, but often clearer | | Skip AI | You’re making a high-stakes decision and can’t independently verify the evidence | More work now; less risk of trusting a fabricated or mismatched source |
The useful combination is often an agent for discovery and a dashboard or explorer for confirmation. A sleek demo that returns links without showing how each link supports its claim is mid. For a protocol decision, source quality beats polished formatting.
Common mistakes
- Treating any citation as proof. Open it and find the exact passage, transaction, or data point that supports the claim.
- Confusing a proposal with an active change. Check execution and current state, not just the vote outcome.
- Ignoring chain and timestamp. Record both. A correct value on Ethereum may be wrong for Arbitrum, and a correct value yesterday may be stale today.
- Letting confidence replace evidence. Ask for “unverified” where support is missing. Don’t fill gaps with an agent’s explanation.
FAQ
Can I trust an AI agent for DeFi research?
Trust it to help find and organize evidence, then verify consequential claims against primary sources. If it can’t provide sources you can open and inspect, treat its answer as a lead, not a conclusion.
What sources are best for checking a DeFi protocol claim?
Use the source that can prove the specific claim: an execution transaction or contract state for an onchain change, an official dashboard for current metrics, and governance records for proposal history. Cross-check important details because each source can be incomplete or delayed.
How do I spot a hallucinated citation?
Open the link and confirm the page exists, refers to the right protocol and chain, and contains evidence for the exact claim. A real link that doesn’t support the statement is still a bad citation.
→ Ask the index what to build your crypto research stack
→ Free credits for these tools
Written by McKlaud AI. Want to know which AI tools actually fit your business? Get a free AI audit.