Verification & Trust
Validate AI outputs with citation checks, cross-source verification, and trust scoring.
Overview
Praxiom AI includes a verification system that validates the quality and accuracy of AI-generated outputs. Every agent response can be checked for citation coverage, cross-source consistency, severity justification, and evidence density. Results appear as trust badges on messages with expandable details.
How It Works
Verification runs in two phases:
Synchronous Verification
Performed during the stream and included in the validation_sync SSE event. Checks that run quickly enough to include in real time:
- Citation coverage — Are claims backed by specific research sources?
- Quote accuracy — Do cited quotes actually appear in the source material?
Asynchronous Verification
Deeper checks that run after the stream completes. Fetched via GET /api/messages/{message_id}/validation:
- Cross-source validation — Do multiple independent sources confirm the same finding?
- Severity analysis — Are severity ratings justified by the evidence weight?
- Density scoring — How evidence-dense is the response relative to its length?
- Diversity check — Does the response draw from a variety of sources or lean on a single one?
Trust Badge
Each verified message displays a trust badge with one of three states:
| Status | Meaning |
|---|---|
| Pass | All checks passed — high confidence in accuracy |
| Warn | Some checks raised concerns — review flagged items |
| Fail | Critical issues found — claims may not be well-supported |
Verification Details
Expanding the trust badge reveals individual check results:
| Check | What It Validates |
|---|---|
citation_coverage | Percentage of claims backed by source citations |
cross_source | Whether findings are corroborated across multiple sources |
severity_justified | Whether severity ratings match the evidence weight |
density | Evidence-to-content ratio |
diversity | Source diversity — how many distinct sources are referenced |
Each check includes:
- Status — pass, warn, or fail
- Score — Numeric confidence (0-1)
- Details — Explanation of what was found
- Sub-checks — Granular per-citation results with similarity scores
Citation Details
For citation-related checks, the verification panel shows:
- The exact quote or claim
- The source it's attributed to
- A similarity score indicating how closely the claim matches the source text
- The source passage for comparison
When Verification Runs
Verification is most valuable after synthesis and recommendation operations where accuracy directly impacts product decisions. It runs automatically for:
- Synthesis outputs (insight extraction)
- Recommendation generation
- Document drafting with citations
For general chat responses, verification is available on-demand.
Use verification results to guide your triage. If a synthesis output shows low cross-source validation, it may indicate the finding is based on a single data point and needs more evidence before acting on it.
Editor Intelligence
The document editor includes a related Intelligence Panel with five tabs:
| Tab | Description |
|---|---|
| Proofread | Grammar, clarity, and consistency checks |
| Verify | Citation and evidence verification |
| Citations | Source attribution and quote accuracy |
| Gaps | Missing sections, unanswered questions |
| Summary | Auto-generated executive summary |
What's Next
You've completed the Praxiom AI feature guides. Explore the API Reference for detailed endpoint documentation, or check out the Tutorials for hands-on walkthroughs.
Was this helpful?