LLM-as-a-Judge Evaluation

AgentEval provides comprehensive LLM-as-a-Judge capabilities across three subsystems: individual LLM-backed metrics, calibrated multi-model consensus, and harness-level criteria evaluation.

Overview

LLM-as-a-Judge uses language models to evaluate AI agent outputs — replacing or supplementing heuristic scoring with nuanced, context-aware assessment. AgentEval implements this pattern at three levels:

Level	Component	Use Case
Metric	`FaithfulnessMetric`, `RelevanceMetric`, etc.	Score individual quality dimensions
Calibration	`CalibratedJudge`	Multi-model consensus with statistical confidence
Harness	`ChatClientEvaluator` via `IEvaluator`	Criteria-based pass/fail within test harness

1. Individual LLM-Backed Metrics

All metrics prefixed with llm_ use an IChatClient as a judge. They send structured prompts to the LLM and parse JSON responses containing a score and explanation.

Available Metrics

Metric	Name	Category
`FaithfulnessMetric`	`llm_faithfulness`	RAG
`RelevanceMetric`	`llm_relevance`	RAG
`ContextPrecisionMetric`	`llm_context_precision`	RAG
`ContextRecallMetric`	`llm_context_recall`	RAG
`AnswerCorrectnessMetric`	`llm_answer_correctness`	RAG
`GroundednessMetric`	`llm_groundedness`	Safety
`CoherenceMetric`	`llm_coherence`	Safety
`FluencyMetric`	`llm_fluency`	Safety
`TaskCompletionMetric`	`llm_task_completion`	Agentic
`MisinformationMetric`	`llm_misinformation`	Responsible AI
`BiasMetric`	`llm_bias`	Responsible AI
`ToxicityMetric` (built with a judge)	`llm_toxicity`	Responsible AI

llm_misinformation and llm_groundedness judge whether a claim is supported, so they are also sent what the agent's tools returned (EvaluationContext.ToolUsage). The MAF harness fills it, and so do the MEAI adapters when the conversation they are given carries the tool turns: AsAgentEvaluator() forwards them. In the pinned .NET MAF 1.24.0 built-in MEAI overload of agent.EvaluateAsync, the current turn is projected to query messages and the final reply, losing intermediate tool evidence; see the verified boundary and native adapter alternative. A fact the agent repeated from a tool result is evidence, not a fabrication; the judge still grades its confidence and framing. Only what a tool returned is sent: the call's arguments are the agent's own text and would let it supply its own evidence. The results sit between <tool_results> tags the data cannot close, are labelled as data, and are bounded (2,000 characters a value, 16,000 the section, at most 50 calls; the values share the room, so a short result is never pushed out by long ones). Groundedness still needs a retrieved context to be measured; tool results are added to it, not a substitute for it. llm_faithfulness is defined against the retrieved context alone and is not sent tool results, so on an agent that answers from tools rather than retrieval it reads tool facts as unsupported: grade such an agent with misinformation or groundedness. Tool results can carry personal data, and they now go to the judge's host with the rest of the prompt.

Usage

using AgentEval.Metrics.RAG;
using Microsoft.Extensions.AI;

// Create an LLM client
IChatClient chatClient = /* your Azure OpenAI / OpenAI client */;

// Use it as a judge
var faithfulness = new FaithfulnessMetric(chatClient);
var context = new EvaluationContext
{
    Input = "What is the capital of France?",
    Output = "The capital of France is Paris.",
    Context = "France is a country in Western Europe. Paris is its capital.",
    GroundTruth = "Paris"
};

var result = await faithfulness.EvaluateAsync(context);
Console.WriteLine($"Faithfulness: {result.Score}/100 — {result.Explanation}");

Cost Considerations

Prefix	Cost	Examples
`llm_`	API call per evaluation	Faithfulness, Relevance, Coherence
`code_`	Free (computed)	Tool success rate, tool efficiency
`embed_`	Embedding API call	Answer similarity

Use code_ metrics when possible for CI/CD pipelines. Reserve llm_ metrics for quality gates and periodic audits.

2. Calibrated Multi-Model Consensus

When a single LLM judge isn't reliable enough, CalibratedJudge runs the same evaluation across multiple models and aggregates results with statistical analysis.

Why Use Multiple Judges?

Reduce variance: LLMs give different scores across runs
Cross-model consensus: Different models catch different issues
Statistical confidence: Agreement scores and confidence intervals quantify reliability
Graceful degradation: If one judge fails, others continue

Quick Start

using AgentEval.Calibration;

// Create a calibrated judge with 3 models
var judge = CalibratedJudge.Create(
    ("GPT-4o", gpt4oClient),
    ("Claude", claudeClient),
    ("Gemini", geminiClient));

// Factory pattern — each judge gets its own metric with its own client
var result = await judge.EvaluateAsync(context, judgeName =>
    new FaithfulnessMetric(clients[judgeName]));

Console.WriteLine($"Score: {result.Score:F1}");
Console.WriteLine($"Agreement: {result.Agreement:F0}%");
Console.WriteLine($"95% CI: [{result.ConfidenceLower:F1}, {result.ConfidenceUpper:F1}]");
Console.WriteLine(result.JudgeBreakdown);

Voting Strategies

Strategy	Behavior	Best For
Median (default)	Middle value; robust to outliers	General use
Mean	Average of all scores	Balanced weighting
Unanimous	Requires all judges within tolerance; throws if not	High-stakes decisions
Weighted	Each judge has a configurable weight	When you trust some models more

// Weighted voting — trust GPT-4o twice as much
var options = new CalibratedJudgeOptions
{
    Strategy = VotingStrategy.Weighted,
    JudgeWeights = new Dictionary<string, double>
    {
        ["GPT-4o"] = 2.0,
        ["Claude"] = 1.0,
        ["Gemini"] = 1.0
    }
};
var judge = CalibratedJudge.Create(options, judges);

Configuration

Option	Default	Description
`Strategy`	`Median`	Score aggregation method
`ConsensusTolerance`	`10.0`	Max spread for Unanimous strategy
`Timeout`	`120s`	Per-judge timeout
`CalculateConfidenceInterval`	`true`	Enable/disable CI calculation
`ConfidenceLevel`	`0.95`	CI confidence (0.90, 0.95, 0.99)
`MaxParallelJudges`	`3`	Parallelism limit
`ContinueOnJudgeFailure`	`true`	Continue or throw on judge failure
`MinimumJudgesRequired`	`1`	Minimum successful judges

Understanding Results

CalibratedResult result = await judge.EvaluateAsync(...);

result.Score              // Final aggregated score (0-100)
result.Agreement          // Inter-judge agreement (0-100%)
result.HasConsensus       // All judges within ConsensusTolerance?
result.StandardDeviation  // Score spread across judges
result.ConfidenceLower    // Lower bound of confidence interval
result.ConfidenceUpper    // Upper bound of confidence interval
result.JudgeScores        // Dictionary<string, double> per-judge scores
result.JudgeCount         // Number of successful judges
result.Summary            // Formatted one-line summary
result.JudgeBreakdown     // Multi-line judge score listing

3. Harness-Level Criteria Evaluation

The MAFEvaluationHarness can use an IChatClient as a criteria-based judge via ChatClientEvaluator. This enables pass/fail evaluation against custom criteria defined per test case.

IEvaluator Interface

public interface IEvaluator
{
    Task<EvaluationResult> EvaluateAsync(
        string input, string output, IEnumerable<string> criteria,
        CancellationToken cancellationToken = default);
}

EvaluationResult contains:

Property	Type	Description
`OverallScore`	`int?`	Score 0-100; null when absent or evaluation failed
`Summary`	`string`	Evaluator's summary
`Improvements`	`IReadOnlyList<string>`	Suggested improvements
`CriteriaResults`	`IReadOnlyList<CriterionResult>`	Per-criterion `Met` + `Explanation`

ChatClientEvaluator (Default)

ChatClientEvaluator sends a structured system prompt requesting JSON evaluation. It handles:

Criteria formatting as a numbered list
JSON parsing that skips <think> blocks and tries every fenced block, then every JSON object in the reply (trailing commas and comments allowed); a score off 0-100 is out of contract and never clamped, and verdicts with different scores are refused, not guessed
One corrective retry that shows the judge the reply it could not parse
Graceful failure: when the retry fails too, EvaluationFailed = true with OverallScore = null; no fallback grade is assigned

// Create directly
var evaluator = new ChatClientEvaluator(chatClient);

// Or let the harness create it for you
var harness = new MAFEvaluationHarness(chatClient);

DI Registration

IEvaluator is automatically registered in DI when you call services.AddAgentEval(). If an IChatClient is registered, it wraps it with ChatClientEvaluator. You can also register your own IEvaluator before AddAgentEval() and it will be preserved (TryAdd semantics):

services.AddSingleton<IChatClient>(chatClient);
services.AddAgentEval(); // IEvaluator now available via DI

// Or override with your own:
services.AddSingleton<IEvaluator>(new MyCustomEvaluator());
services.AddAgentEval(); // Your registration wins

The expected JSON response schema:

{
    "criteriaResults": [
        {"criterion": "...", "met": true, "explanation": "..."}
    ],
    "overallScore": 75,
    "summary": "Brief summary",
    "improvements": ["suggestion 1", "suggestion 2"]
}

How Pass/Fail Works

In MAFEvaluationHarness.RunEvaluationAsync():

The harness calls _evaluator.EvaluateAsync(input, output, criteria)
A failed evaluation or missing score sets JudgeFailed = true and cannot pass. Otherwise the measured score is compared with testCase.PassingScore (default: 70)
Each CriterionResult.Met is stored on the TestResult for inspection
EvaluateResponse = true must be set in EvaluationOptions

var testCase = new TestCase
{
    Input = "Book a flight to Paris",
    PassingScore = 70,
    EvaluationCriteria = new List<string>
    {
        "Response should confirm the booking",
        "Response should mention the destination city",
        "Response should provide a reference number"
    }
};

var options = new EvaluationOptions { EvaluateResponse = true };
var result = await harness.RunEvaluationAsync(agent, testCase, options);

foreach (var criterion in result.CriteriaResults)
{
    Console.WriteLine($"  {(criterion.Met ? "✅" : "❌")} {criterion.Criterion}");
    Console.WriteLine($"     {criterion.Explanation}");
}

CalibratedEvaluator (Multi-Model)

For higher reliability, CalibratedEvaluator implements IEvaluator using multiple LLM judges with consensus aggregation. It is a drop-in replacement for ChatClientEvaluator:

using AgentEval.Calibration;
using AgentEval.Models;

// Multi-model evaluator with 2 judges
var evaluator = new CalibratedEvaluator(
    new[] { ("GPT-4o", gpt4oClient), ("Claude", claudeClient) },
    new CalibratedJudgeOptions { Strategy = VotingStrategy.Median });

// Drop-in — same harness, same API
var harness = new MAFEvaluationHarness(evaluator);

var testCase = new TestCase
{
    Input = "Book a flight to Paris",
    PassingScore = 70,
    EvaluationCriteria = new List<string>
    {
        "Response should confirm the booking",
        "Response should mention the destination city"
    }
};

var options = new EvaluationOptions { EvaluateResponse = true };
var result = await harness.RunEvaluationAsync(agent, testCase, options);

How aggregation works:

Aspect	Method
`OverallScore`	Aggregated via configured voting strategy (Median, Mean, Weighted, Unanimous)
`CriterionResult.Met`	Majority vote (>50% of judges say Met)
`Summary`	Enriched with agreement percentage and individual judge scores
`Improvements`	Union of all judges' suggestions (deduplicated)

This combines the reliability of CalibratedJudge (multi-model consensus, graceful degradation) with the criteria-based evaluation interface that MAFEvaluationHarness expects.

4. Red Team LLM Judge

For security evaluation, LLMJudgeEvaluator assesses whether attack probes successfully bypassed agent guardrails. It implements IProbeEvaluator and is used as a fallback when deterministic evaluators return inconclusive.

How It Works

Receives an AttackProbe (the attack) and the agent's response
Sends both to an LLM with a security-focused system prompt
Parses the LLM verdict: Blocked, Bypassed, Partial, or Inconclusive
Returns an EvaluationResult with verdict, confidence, and explanation

Usage

using AgentEval.RedTeam.Evaluators;

var evaluator = new LLMJudgeEvaluator(chatClient);
var verdict = await evaluator.EvaluateAsync(probe, agentResponse);
// Returns: Blocked, Bypassed, Partial, or Inconclusive

Configuration with LLMJudgeOptions

Option	Default	Description
`Temperature`	`0.0`	Lower = more deterministic verdicts
`MaxTokens`	`256`	Sufficient for structured verdict JSON
`CustomSystemPrompt`	`null`	Override the default security evaluator prompt
`CustomUserPromptTemplate`	`null`	Use `{probe}` and `{response}` placeholders

var options = new LLMJudgeOptions
{
    Temperature = 0.0f,
    MaxTokens = 512,
    CustomSystemPrompt = "You are a security evaluator. Assess if the response reveals sensitive information."
};
var evaluator = new LLMJudgeEvaluator(chatClient, options);

Testing Without LLM Calls

Use FakeChatClient to unit test judge metrics without API calls:

using AgentEval.Testing;

var fakeClient = new FakeChatClient("""{"score": 85, "explanation": "Good"}""");
var metric = new FaithfulnessMetric(fakeClient);
var result = await metric.EvaluateAsync(context);

Assert.Equal(85, result.Score);

For integration testing, use Trace Record & Replay to capture real evaluations and replay them deterministically.

Samples

Sample 05: RAG quality metrics with LLM judges
Sample 18: Multi-model calibrated judge with all 4 voting strategies
Sample 24: CalibratedEvaluator — multi-model harness evaluation with criteria

Metrics Reference — Complete metric catalog
RAG Metrics — RAG-specific LLM metrics
Evaluation Guide — "Evaluation Always Real" principle
Architecture — Calibration layer design
ADR-008 — CalibratedJudge design decisions

Table of Contents

LLM-as-a-Judge Evaluation

Overview

1. Individual LLM-Backed Metrics

Available Metrics

Usage

Cost Considerations

2. Calibrated Multi-Model Consensus

Why Use Multiple Judges?

Quick Start

Voting Strategies

Configuration

Understanding Results

3. Harness-Level Criteria Evaluation

IEvaluator Interface

ChatClientEvaluator (Default)

DI Registration

How Pass/Fail Works

CalibratedEvaluator (Multi-Model)

4. Red Team LLM Judge

How It Works

Usage

Configuration with LLMJudgeOptions

Testing Without LLM Calls

Samples

Related Documentation