Evaluation Framework

The Evaluation Framework provides comprehensive assessment of AI outputs across multiple evaluation dimensions. Instead of relying on a single metric, it runs multiple independent evaluations and uses statistical consensus to reduce variance and improve reliability.

Available Evaluators

TruthVouch includes 9 built-in evaluators covering diverse quality dimensions:

• Factual Accuracy — Checks if claims are factually correct against your knowledge base
• Semantic Similarity — Measures whether the output captures the intended meaning
• Response Completeness — Ensures the response fully addresses the user’s request
• Hallucination Score — Detects fabricated information not grounded in source material
• Citation Verification — Confirms citations accurately reference sources
• Prompt Injection Detection — Identifies adversarial prompts attempting to override system behavior
• Toxicity & Safety — Flags harmful, offensive, or unsafe content
• Bias Detection — Identifies stereotyping, unfair representation, or discriminatory language
• Custom Evaluators — Define your own evaluation criteria and scoring rubrics

ChainPoll Consensus

ChainPoll runs multiple LLM-based evaluations on the same output and combines results using statistical consensus:

Input → [Evaluator A] → Score: 0.92
         [Evaluator B] → Score: 0.88
         [Evaluator C] → Score: 0.90
         ↓
      Consensus: 0.90 (mean)
      Confidence: 0.85 (low variance)

This approach reduces individual evaluator variance and improves reliability. The consensus score represents the mean across all runs, while confidence reflects the variance:

• High variance (low confidence) — Evaluators disagree; result is uncertain
• Low variance (high confidence) — Evaluators agree; result is reliable

Configuration

When running an evaluation, specify:

• Number of samples — How many times to run each evaluator (default: 3)
• Consensus threshold — Minimum agreement level required to pass (default: 0.75)
• Evaluator selection — Which evaluators to run (run all, or select specific ones)

Custom Evaluator Builder

Create custom evaluators tailored to your domain:

1. Go to Governance Hub → Evaluation Framework
2. Click Create Custom Evaluator
3. Define:
   • Name — Descriptive evaluator name (e.g., “Medical Terminology Accuracy”)
   • Description — What this evaluator measures
   • Rubric — Scoring criteria (1-5 point scale, with descriptors for each level)
   • Examples — Sample inputs and expected scores for calibration
4. Test the evaluator against sample outputs
5. Deploy to make available for all evaluations

Custom evaluators are evaluated using the same LLM-based approach as built-in evaluators, allowing them to learn from your examples.

Agentic Evaluation Metrics

For agentic AI systems (agents that take actions, call tools, etc.), specialized metrics assess:

• Tool Selection Accuracy — Did the agent choose the right tool for the task?
• Action Completion Rate — What percentage of tool calls succeeded?
• Tool Error Recovery — When a tool fails, does the agent retry or recover gracefully?
• Plan Coherence — Does the agent’s action sequence make logical sense?
• Resource Efficiency — Did the agent accomplish the task with minimal tool calls?

These metrics are automatically available when evaluating agent outputs.

Uncertainty Scoring

Evaluation results include uncertainty estimates computed via multi-sample variance analysis:

• Confidence Score (0-1) — How confident the evaluation result is
• Variance — Statistical spread across samples
• Sample Count — Number of evaluations run

Example:

{
  "evaluator": "Factual Accuracy",
  "score": 0.87,
  "confidence": 0.92,
  "variance": 0.015,
  "samples": 5,
  "interpretation": "High confidence: evaluators consistently agree"
}

Use confidence scores to:

• Flag uncertain results for manual review (confidence < 0.7)
• Require additional sampling on borderline scores (0.4-0.6)
• Automate decisions on high-confidence results (> 0.85)

Evaluation Framework UI

The Evaluation Framework lives in the Governance Hub and provides:

Configuration Screen

• Select evaluators to run
• Set consensus thresholds
• Configure sample counts
• Manage custom evaluators
• View evaluator performance history

Run Evaluations

Input Batch → Configure Evaluators → Run → Monitor Progress
                                    ↓
                            Results Dashboard

Results Dashboard

View evaluation results across dimensions:

• Evaluator Scores — Table showing each evaluator’s score and confidence
• Consensus Score — Overall evaluation result
• Trend Analysis — How scores change over time
• Failed Evaluations — Which specific checks failed and why
• Distribution — Histogram of scores across your evaluation history

Custom Evaluator Management

• List all custom evaluators with performance metrics
• Edit rubrics and examples
• Test evaluator against sample data
• Archive or delete unused evaluators
• Review evaluator training examples

Integration Examples

Batch Evaluation

Evaluate multiple outputs at once:

from truthvouch import TruthVouchClient

client = TruthVouchClient(api_key="your-api-key")

outputs = [
  "The capital of France is Paris.",
  "Machine learning is a type of artificial intelligence.",
  "The moon orbits the earth in 28 days."
]

results = []
for output in outputs:
    result = client.evaluate_output(
        text=output,
        model="gpt-4"
    )
    results.append({
        "output": output,
        "blocked": result.blocked,
        "flagged": result.flagged
    })

Streaming Evaluation

For streaming responses, evaluate in chunks or at completion:

# Evaluate complete response after streaming
full_response = ""
async for chunk in llm.stream_response(prompt):
    full_response += chunk

# Final evaluation
result = await client.evaluate_output(
    text=full_response,
    model="gpt-4"
)

if result.blocked:
    # Response violates policy
    return {"error": result.block_reasons}

Best Practices

Choosing Evaluators

• General content — Use Hallucination, Toxicity & Safety, Bias Detection
• Domain-specific — Add custom evaluators for your industry
• Agent outputs — Use agentic metrics for agent evaluation
• Strict compliance — Run all evaluators; set high consensus thresholds

Setting Thresholds

• Critical content (medical, legal, financial) — 0.9+ consensus, all evaluators
• General use — 0.75 consensus, standard evaluators
• Uncertain cases — Flag for human review if confidence < 0.7

Sampling Strategy

• Fast evaluation — 1-2 samples, trade off reliability for speed
• Balanced — 3-5 samples (default), good reliability with acceptable latency
• High assurance — 10+ samples, maximum reliability for critical decisions

Next Steps

• Create your first custom evaluator
• Configure evaluation policies
• Monitor evaluation metrics in dashboards