Understanding Results

Learn how to interpret optimization results and apply improvements.

Results Overview

When an optimization completes, you'll see:

• Winner: The best-performing variant (or original if none improved)
• Lift: Percentage improvement over the original
• Confidence: How reliable the result is
• Metrics: Detailed performance breakdown

Reading the Results

Winner Status

Status	Meaning
Variant Won	A variant outperformed the original
Original Won	Your agent is already performing well
Inconclusive	Not enough difference to declare a winner

Lift Percentage

The improvement compared to your original prompt:

• +20% task completion = Users complete their goals 20% more often
• +15% sentiment = Users feel 15% more positive about interactions
• -10% response length = Responses are 10% more concise

Confidence Level

How sure we are about the result:

Confidence	Meaning
High (>95%)	Very reliable, safe to deploy
Medium (80-95%)	Likely accurate, consider more testing
Low (<80%)	Inconclusive, run more simulations

Viewing Variant Details

Dashboard

Click on any variant to see:

• Full prompt content
• Side-by-side comparison with original
• Sample conversations
• Metric breakdown

Via MCP

Show me the variants from my last optimization

Compare variant B to my original prompt

Via SDK

const variants = await converra.optimizations.getVariants('opt_123');

variants.forEach(v => {
  console.log(`${v.name}:`);
  console.log(`  Task completion: ${v.metrics.taskCompletion}%`);
  console.log(`  Lift: ${v.metrics.lift}%`);
});

Metrics Explained

Task Completion

Did the AI help users achieve their goal?

• High: Users got what they needed
• Low: Users left without resolution

Response Quality

Was the response accurate, helpful, and appropriate?

• High: Clear, correct, actionable responses
• Low: Vague, incorrect, or unhelpful

User Sentiment

How would users feel about the interaction?

• Positive: Satisfied, happy
• Neutral: Neither satisfied nor dissatisfied
• Negative: Frustrated, disappointed

Conciseness

Are responses appropriately sized?

• Good: Right length for the context
• Too long: Verbose, could be shortened
• Too short: Missing important information

Testing Before Applying

Want to validate the winner before deploying? Use Test Winner to run additional simulations.

Dashboard

From the optimization dropdown menu, select Test Winner. This runs the winning variant through another round of simulations to confirm performance.

Via MCP

Test the winner from my last optimization

When to Use

• High-stakes prompts where you want extra confidence
• When the lift is marginal and you want to verify
• Before deploying to production after a long optimization

Deploying the Winner

Deployment works like CI/CD for agent improvements. Every fix passes simulation and regression testing before it ships — choose how it lands:

Auto-deploy

Enable auto-deploy in Settings and winners ship automatically when they pass regression testing. If any metric regresses after deployment, it rolls back before the next customer conversation.

GitHub PR

Converra creates a PR with the improved prompt, metrics, evidence, and check runs. Merge to deploy. Works like Dependabot for AI agents.

See GitHub Integration for setup.

Dashboard / MCP / SDK

Apply manually when you prefer:

Apply the winning variant from my last optimization

await converra.optimizations.applyVariant('opt_123');

What Happens

1. Your agent's prompt content is updated to the winning variant
2. A new version is created in version history
3. Your original is preserved (can revert anytime)
4. If any metric regresses, it rolls back automatically
5. Cache is invalidated (SDK users get new content immediately)

Regression Test Results

When a variant shows improvement, Converra automatically runs regression tests. Results appear alongside winner metrics:

Regression Check: PASSED (5/5 scenarios)

or

Regression Check: 1 REGRESSION

4/5 scenarios passed
✗ Technical support inquiry: -16%

Apply anyway? [Yes] [No - Keep Baseline]

Regressions don't automatically block deployment—you see the tradeoff and decide.

See Regression Testing for details.

When No Clear Winner

If results are inconclusive:

1. Run validation mode - More simulations = clearer results
2. Adjust intent - Focus on specific improvements
3. Review manually - Sometimes human judgment is needed
4. Keep original - If it's working, don't change it

Learning from Results

What Worked

Look at winning variant changes:

• Added examples? → Examples help.
• Restructured? → Format matters.
• Changed tone? → Audience preference revealed.

What Didn't Work

Failed variants show what to avoid:

• Too formal? Too casual?
• Too verbose? Too terse?
• Missing context? Over-explained?

Sample Results

Optimization Complete: opt_abc123

Winner: Variant B (+23% lift)
Confidence: High (97%)

Metrics vs Original:
┌──────────────────┬──────────┬───────────┬────────┐
│ Metric           │ Original │ Variant B │ Change │
├──────────────────┼──────────┼───────────┼────────┤
│ Task Completion  │ 72%      │ 89%       │ +17%   │
│ Response Quality │ 81%      │ 94%       │ +13%   │
│ User Sentiment   │ 68%      │ 85%       │ +17%   │
│ Conciseness      │ 65%      │ 78%       │ +13%   │
└──────────────────┴──────────┴───────────┴────────┘

Key Changes in Variant B:
- Added step-by-step format for instructions
- Included acknowledgment before solutions
- Added follow-up confirmation question

Next Steps

• Logging Conversations - Track real performance
• Analyzing Insights - Understand patterns