Docs/API Reference/Verification

Claim Verification

Every Alphagent response includes a verification report that validates claims against source data. This helps you trust the information and identify potential issues.

Built-in Verification

All models include a Verifier agent that automatically checks claims against tool outputs. No additional configuration required.

#Accessing Verification

The verification report is included in the response metadata:

Python

response = client.responses.create(
    model="alphagent-deep-research-pro",
    input="Analyze AAPL's current valuation"
)

# Access verification report
if response.metadata:
    verification = response.metadata.get("verification", )
    print("Confidence:", verification.get("confidence"))
    print("Verified claims:", verification.get("verified_claims"))
    print("Flagged claims:", verification.get("flagged_claims"))

#Verification Fields

verified_claims

Array of strings describing claims that were successfully matched to supporting tool output. These claims have been validated against real data sources.

flagged_claims

Array of objects with claim, issue, and severity. These claims could not be fully verified or have potential issues.

confidence

Overall confidence level: "high", "medium", or "low". Based on the ratio of verified to flagged claims.

#Severity Levels

Flagged claims include a severity level to help you prioritize:

Severity	Meaning
info	Claim couldn't be verified but isn't necessarily wrong. May lack supporting data.
warning	Claim may be inaccurate or outdated. Should be verified manually.
error	Claim contradicts available data. Likely incorrect.

#Example Flagged Claim

Here's what a flagged claim looks like in the response:

JSON

{
  "claim": "Netflix shares slid 5-6%",
  "issue": "No price tool was called for NFLX to verify the specific percentage drop.",
  "severity": "warning"
}

#Best Practices

Check confidence for critical decisions

For financial decisions, prefer responses with "high" confidence. If confidence is "low", consider re-querying or manually verifying the information.

Display flagged claims to users

If building user-facing applications, consider showing flagged claims with appropriate warnings so users know which parts of the response may need verification.

Use deep research models for complex queries

Multi-agent models (alphagent-deep-research-*) provide more thorough verification since they call more tools and cross-reference more sources.

Streaming