Analytics
Analytics gives you the portfolio view. Use it when you want trend reporting, comparisons, cost visibility, and model-level analysis instead of a deep dive into one session.
Best for:
- Tracking policy score, risk, and alert trends over time.
- Comparing two cohorts before and after a model, prompt, or rule change.
- Finding which applications, models, and intents drive the most LLM spend.
- Asking grounded questions about fleet-wide performance.
Exports: Every chart can be exported as CSV or PNG from the chart toolbar.
Date range: The global date-range picker controls every chart on the page.
Tip: Enable Technical view to add latency, payload, and enrichment charts for deeper operational diagnostics.
Analytics
The Analytics workspace helps you answer the questions that matter when you run AI systems at scale: what is improving, what is drifting, where cost is concentrating, and which changes affected quality or policy outcomes.
Analytics is role-gated: Administrators see all tabs. Governance Engineers see the Trends, Portfolio Q&A, Cohort Comparison, and Annotation Analytics tabs. Business Owners see the Trends, Portfolio Q&A, Cohort Comparison, Cost, Models, and Financial tabs. Developers, Compliance Officers, and Auditors do not have Analytics page access — their trend data surfaces in the Compliance workspace and Governance Dashboard instead.
Your Analytics view includes the Trends, Portfolio Q&A, and Cohort Comparison base tabs plus the Annotation Analytics tab. Use Annotation Analytics to track annotation coverage rates, exemplar quality trends, and labeling distribution across your application fleet — so you know whether your evaluation datasets are keeping pace with application volume.
Your Analytics view includes the Trends, Portfolio Q&A, and Cohort Comparison base tabs plus the Cost, Models, and Financial tabs. Use Cost to track LLM spend attribution by application and model. Use Trends to watch how risk distribution and session volume change over time. Together they give you the data you need to justify AI investment and identify where cost is concentrating.
You see every Analytics tab: Trends, Portfolio Q&A, Cohort Comparison, Cost, Models, Financial, and Annotation Analytics. Use the Technical view toggle to add latency, payload size, and enrichment-rate charts for deeper operational diagnostics.
Trends
The Trends view is the default tab. It shows the main policy and operational signals across your tenant.
| Chart | Description |
|---|---|
| Policy score | Rolling average per application, shown for all applications or any selected subset |
| Risk distribution | Area chart showing the mix of LOW / MEDIUM / HIGH / CRITICAL sessions over time |
| Session volume | Daily ingestion count per application |
| Alert and guardrail activity | Daily count of blocked, flagged, allowed, and alert-triggering events |
| Human oversight coverage | Percentage of sessions with a human oversight annotation, trended over time |
| Anchor coverage | Blockchain anchor success rate over time |
Date range selector: Choose from 7-day, 30-day, 90-day, or a custom range. All charts update together.
Application selector: Review all applications together or isolate specific ones. Selecting multiple applications overlays trend lines for direct comparison.
Start your weekly policy review here. If a score, risk mix, or anchor rate changes sharply, click the chart point and move into the Sessions for investigation.
Cohort Comparison
The Cohort view compares two groups of sessions side by side. Build each cohort from any combination of:
- Date range
- Application(s)
- Model
- Environment (Production / Staging)
- Intent label (partial match)
The comparison shows score, risk, outcome, and guardrail differences between the two groups. Use it to answer questions like:
- Did policy performance improve after the March 3 model update?
- Is the staging environment behaving differently from production?
- Are sessions with intent “loan-decision” riskier than “account-inquiry”?
Cohort results can be exported to CSV for review decks, audit preparation, or BI analysis.
Cost Attribution
The Cost view tracks estimated LLM spend across your tenant by application, model, and intent. It uses SDK token counts and the pricing table you configure in Settings → Billing → Model Pricing.
| Panel | Description |
|---|---|
| Total estimated cost | Aggregate LLM API spend for the selected period |
| Cost by application | Bar chart showing spend attribution per application |
| Cost by model | Pie chart showing which models are driving the most spend |
| Cost by intent | Breakdown of spend associated with each intent category |
| Cost trend | Daily cost trend with budget overlay (if a cost budget is configured) |
Cost figures are estimates. They reflect the token and pricing data configured in VeriProof, not your provider invoice, so they should not be used for formal financial reporting.
Models
The Models view compares quality and policy metrics across every model identifier found in your session data. For each model, it shows:
| Metric | Description |
|---|---|
| Session count | Total sessions using this model in the selected period |
| Average policy score | Mean policy score across all sessions |
| Risk distribution | Proportion of sessions at each risk level |
| Average confidence | Mean model confidence score |
| Guardrail hit rate | Percentage of sessions where a guardrail evaluated this model’s output |
| Guardrail block rate | Percentage of all sessions where a guardrail blocked this model’s output |
Use this view during a model upgrade or rollback review. Compare the new model against the old one inside the same application and intent scope.
Portfolio Q&A
Portfolio Q&A accepts natural-language questions about your tenant data and returns grounded answers with supporting figures.
Example queries:
- “Which application had the most HIGH-risk decisions in Q1 2026?”
- “Show me the week-over-week governance score change for the Loan Assistant application.”
- “How many sessions were blocked by guardrails in the last 30 days?”
Answers cite the data used to reach the conclusion. When a chart would help, it is rendered inline with the response.