metric-design

Metric Design

Select, configure, and combine evaluation graders into a working pipeline. You choose the right tool for each evaluation dimension — from zero-cost code checks to LLM judges — and produce executable GradingRunner code that runs on OpenJudge.

Requires OpenJudge (pip install py-openjudge). This skill is intentionally SDK-centric — grader selection, GradingRunner, and aggregators are OpenJudge APIs. The design/decision logic still applies if you use another harness; only the code does not.

When to Activate

• User has eval dimensions/principles but doesn't know which grader type to use
• User wants to write an LLM-as-judge prompt for a specific failure mode
• User needs a composite score combining multiple evaluation dimensions
• User wants to auto-generate graders from labeled data instead of writing them manually
• User's current evaluation is all LLM-based and too expensive/too slow

Checklist

You MUST create a task for each item and complete them in order:

1. Select grader types — per dimension, pick the right grader class
2. Create custom graders — write judge prompts (4-component) or function graders
3. Auto-generate if applicable — use OpenJudge Generator for cold starts
4. Run anti-pattern scan — check for Likert, missing few-shot, vague criteria
5. Build pipeline code — assemble GradingRunner with graders + aggregators

Step 1: Select Grader Type Per Dimension

For each evaluation dimension, walk this decision tree (first match wins):

1. Can a deterministic rule check this?
   → StringMatchGrader / JsonValidatorGrader / FunctionGrader (zero cost, 100% consistent)
   Examples: exact match for classification labels, regex for format checks,
             JSON schema validation, keyword presence/absence

2. Does it require semantic understanding of text quality?
   → LLMGrader with built-in class (low cost, pre-optimized)
   Examples: CorrectnessGrader (factual match), RelevanceGrader (on-topic check),
             HallucinationGrader (faithfulness to context)

3. Does it involve agent behavior (tool calls, planning, memory)?
   → Agent-specific LLMGrader
   Examples: ToolSelectionGrader, TrajectoryAccuracyGrader, MemoryAccuracyGrader

4. Does it involve code execution or syntax?
   → CodeExecutionGrader / SyntaxCheckGrader
   Examples: test case pass rate, syntax validity, code style checks

5. Does it require external tool calls to verify (web search, database lookup)?
   → AgenticGrader (expensive, use only when necessary)
   Examples: fact-checking against live sources, cross-referencing databases

Grader Selection Cheat Sheet

Output type	Recommended grader	Cost
Classification label	StringMatchGrader	Free
JSON structure	JsonValidatorGrader + JsonMatchGrader	Free
Free text correctness	CorrectnessGrader	LLM call
Factual accuracy (grounded)	HallucinationGrader	LLM call
Response relevance	RelevanceGrader	LLM call
Instruction following	InstructionFollowingGrader	LLM call
Tool call selection	ToolSelectionGrader	LLM call
Agent trajectory	TrajectoryAccuracyGrader	LLM call
Code correctness	CodeExecutionGrader	Free
Custom quality check	Custom LLMGrader	LLM call
External fact verification	AgenticGrader	LLM + tool calls

Why this order matters: Every LLM-based grader adds cost, latency, and non-determinism. A StringMatchGrader costs nothing and always gives the same answer. Exhaust deterministic options before reaching for an LLM judge.

Step 2: Create Custom Graders

LLMGrader: The Four-Component Template

When no built-in grader fits, create a custom LLMGrader. Every judge prompt needs exactly these four components (adapted from community best practice):

Component 1 — Task & Criterion: What this judge evaluates. One thing only.

You are evaluating whether a customer support response correctly identifies
and uses the customer's order number from the conversation context.

Component 2 — Binary Pass/Fail Definitions: Concrete, observable conditions.

PASS: The response references the correct order number exactly as it appears
in the context. If multiple orders exist, the response addresses the right one.

FAIL: The response uses a wrong order number, omits the order number when one
was provided, or references an order not present in the context.

Why binary and not Likert? Because two human annotators agree on "pass vs fail" far more often than on "3 vs 4 out of 5." Binary forces a clear decision boundary. If you need severity levels, use multiple binary judges (e.g., "factually wrong" + "dangerously wrong").

Component 3 — Few-Shot Examples: At minimum 1 pass, 1 fail, 1 borderline. The borderline example is the most valuable — it teaches the judge where the boundary is.

Example 1 (PASS):
Context: "Order #12345: shipped May 10"
Response: "Your order #12345 was shipped on May 10 and arrives May 12."
Critique: The response uses the exact order number (#12345) and matches the
ship date from context. No fabrication or omission.
Result: Pass

Example 2 (FAIL):
Context: "Order #12345: shipped May 10"
Response: "Your order #12346 is on its way!"
Critique: The response uses order #12346 but the context only mentions #12345.
This is a fabricated order number, not a typo — #12346 doesn't exist.
Result: Fail

Example 3 (BORDERLINE PASS):
Context: "Orders #12345 (shipped), #12346 (processing)"
Response: "Your recent order has shipped and should arrive soon."
Critique: The response doesn't specify which order, but says "recent order"
which could reasonably refer to either. If the customer only asked about
shipped items, this is fine. If they asked about a specific order, it's
insufficient. Given the generic phrasing, this passes but is weak.
Result: Pass

Component 4 — Structured Output: Force critique before verdict.

{
  "critique": "Detailed assessment referencing specific evidence from the response and context",
  "result": "Pass" or "Fail"
}

Why critique-before-verdict? LLMs that commit to a verdict first anchor on it and rationalize backward. Reasoning first → verdict second produces more accurate judgments (CoT-then-Score AUC ~0.97 vs verdict-first significantly lower).

Complete LLMGrader Code

from openjudge.graders.llm_grader import LLMGrader
from openjudge.graders.schema import GraderMode

order_accuracy_grader = LLMGrader(
    model=model,
    name="order_accuracy",
    mode=GraderMode.POINTWISE,
    template="""
You are evaluating whether a customer support response correctly identifies
and uses the customer's order number from the conversation context.

Context: {context}
Response: {response}

## Pass/Fail Definitions

PASS: The response references the correct order number exactly as it appears
in the context. If multiple orders exist, the response addresses the right one.

FAIL: The response uses a wrong order number, omits the order number when one
was provided, or references an order not present in the context.

## Examples

Example 1 (PASS):
Context: "Order #12345: shipped May 10"
Response: "Your order #12345 was shipped on May 10 and arrives May 12."
Critique: Exact order number match. Ship date matches context. No fabrication.
Result: Pass

Example 2 (FAIL):
Context: "Order #12345: shipped May 10"
Response: "Your order #12346 is on its way!"
Critique: Order #12346 does not exist in context. Fabricated order number.
Result: Fail

Example 3 (BORDERLINE PASS):
Context: "Orders #12345 (shipped), #12346 (processing)"
Response: "Your recent order has shipped and should arrive soon."
Critique: Doesn't specify which order. "Recent order" is ambiguous but not
factually wrong — it acknowledges a shipped order exists.
Result: Pass

## Output Format

Respond in JSON:
{{"critique": "<detailed assessment>", "result": "Pass" or "Fail"}}
""",
)

FunctionGrader: Deterministic Checks

Use when the rule is code-expressible:

from openjudge.graders.function_grader import FunctionGrader
from openjudge.graders.schema import GraderScore, GraderMode

def no_competitor_mention(response: str, competitors: list[str] = None) -> GraderScore:
    """Check that response doesn't mention competitor brands."""
    if competitors is None:
        competitors = ["competitor_a", "competitor_b", "rival_co"]
    mentioned = [c for c in competitors if c.lower() in response.lower()]
    if not mentioned:
        return GraderScore(name="no_competitor", score=1.0, reason="No competitor mentions")
    return GraderScore(
        name="no_competitor", score=0.0,
        reason=f"Mentioned competitors: {', '.join(mentioned)}"
    )

competitor_grader = FunctionGrader(
    func=no_competitor_mention,
    name="no_competitor",
    mode=GraderMode.POINTWISE,
)

Step 3: Auto-Generate Graders (Cold Start)

When you have no rubric but do have a task description or labeled data, use OpenJudge Generators to create graders automatically:

Zero-shot: SimpleRubricsGenerator

from openjudge.generator.simple_rubric.generator import (
    SimpleRubricsGenerator,
    SimpleRubricsGeneratorConfig,
)

config = SimpleRubricsGeneratorConfig(
    grader_name="Customer Support Quality",
    model=model,
    task_description="Customer support chatbot for e-commerce: orders, returns, shipping",
    scenario="Customers asking about order status, return policies, and delivery times",
    min_score=0,
    max_score=1,
)

generator = SimpleRubricsGenerator(config)
grader = await generator.generate(
    dataset=[],
    sample_queries=[
        "Where is my order?",
        "How do I return this item?",
        "When will my package arrive?",
    ],
)
# grader is now a ready-to-use LLMGrader

Data-driven: IterativeRubricsGenerator

Use when you have 20+ labeled examples (query + response + score):

from openjudge.generator.iterative_rubric.generator import (
    IterativeRubricsGenerator,
    IterativePointwiseRubricsGeneratorConfig,
)

config = IterativePointwiseRubricsGeneratorConfig(
    grader_name="E-commerce QA Grader",
    model=model,
    task_description="Evaluate factual answers to e-commerce customer questions",
    min_score=0,
    max_score=1,
    max_epochs=3,
    batch_size=10,
)

train_data = [
    {"query": "What's your return policy?", "response": "30-day returns, free shipping.", "label_score": 1},
    {"query": "What's your return policy?", "response": "We have a policy.", "label_score": 0},
    # ... 20+ examples
]

generator = IterativeRubricsGenerator(config)
grader = await generator.generate(dataset=train_data)

Step 4: Anti-Pattern Scan

Before finalizing, check every LLM-based grader for these issues:

Check	What to look for	Severity
Likert scale	"rate 1-5", "score 1-10", "Likert" in prompt	BLOCKER — replace with binary Pass/Fail
Missing few-shot	No labeled examples in the prompt	BLOCKER — add at least 1 pass + 1 fail + 1 borderline
Holistic criterion	Single judge evaluating 3+ dimensions	WARNING — split into separate graders, one per dimension
Missing output format	No JSON schema specified	BLOCKER — add {{"critique": "...", "result": "Pass"/"Fail"}}
Vague pass/fail	< 20 words or uses "good"/"bad"/"quality"	WARNING — make definitions concrete and observable
Judge = target model	Same model for both roles	BLOCKER — judge and target must be different models

Why blockers matter: A Likert-scale judge with no few-shot examples and a vague criterion produces scores that look precise but can't be reproduced or calibrated. You'll discover this in production when the judge's TPR/TNR is measured — and it's too late.

Step 5: Build Pipeline Code

First record the design as a metric-plan.yaml so it's reusable and reviewable, then implement it as the runner below:

dimensions:
  - {id: order_accuracy, grader: CorrectnessGrader, mode: score, weight: 0.4,
     mapper: {response: response, reference_response: reference_response}}
  - {id: no_pii, grader: FunctionGrader, mode: gate, gate_threshold: 1.0}  # hard requirement
aggregation: GatedWeightedSumAggregator

Assemble everything into a working GradingRunner:

import asyncio
from openjudge.models.openai_chat_model import OpenAIChatModel
from openjudge.graders.common.correctness import CorrectnessGrader
from openjudge.graders.common.relevance import RelevanceGrader
from openjudge.graders.common.hallucination import HallucinationGrader
from openjudge.graders.text.string_match import StringMatchGrader
from openjudge.runner.grading_runner import GradingRunner, GraderConfig
from openjudge.runner.aggregator.weighted_sum_aggregator import WeightedSumAggregator
from openjudge.graders.schema import GraderScore, GraderError

# Judge model (must differ from the model being evaluated).
# OpenAIChatModel reads OPENAI_API_KEY / OPENAI_BASE_URL from the environment when
# not passed explicitly — point them at any OpenAI-compatible endpoint.
#   OpenAI:          OPENAI_API_KEY=sk-...   (no base_url needed)
#   Aliyun DashScope: OPENAI_API_KEY=<dashscope key>
#                     OPENAI_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
model = OpenAIChatModel(model="qwen-plus")  # or "gpt-4o", etc.

dataset = [
    {
        "query": "Where is my order #12345?",
        "response": "Your order #12345 shipped May 10, arriving May 12.",
        "reference_response": "Order #12345: shipped May 10, ETA May 12. Tracking: 1Z999AA10123456784.",
        "context": "Order #12345 | Status: shipped | Date: 2026-05-10 | Tracking: 1Z999AA10123456784",
    },
    # ... more samples
]

runner = GradingRunner(
    grader_configs={
        # LLM-based graders for semantic quality
        "correctness": CorrectnessGrader(model=model),
        "relevance": RelevanceGrader(model=model),
        "hallucination": HallucinationGrader(model=model),
        # Deterministic grader — zero cost
        "format_json": GraderConfig(
            grader=StringMatchGrader(algorithm="substring_match"),
            mapper={"response": "response", "reference_response": "reference_response"},
        ),
    },
    # Combine into a single weighted score
    aggregators=WeightedSumAggregator(
        name="overall",
        weights={
            "correctness": 0.4,
            "relevance": 0.2,
            "hallucination": 0.3,
            "format_json": 0.1,
        },
    ),
    max_concurrency=8,
)

async def main():
    results = await runner.arun(dataset)

    for grader_name, grader_results in results.items():
        scores = [r.score for r in grader_results if isinstance(r, GraderScore)]
        errors = [r for r in grader_results if isinstance(r, GraderError)]
        avg = sum(scores) / len(scores) if scores else 0
        print(f"{grader_name}: avg={avg:.3f}, errors={len(errors)}")

asyncio.run(main())

Weight Design Principle

Don't use equal weights — they're the most arbitrary choice. Weights should reflect:

• Failure prevalence (from trace analysis): if hallucination failures occur 3x more often than relevance failures, weight hallucination higher.
• Business impact: a correctness failure might cost a customer; a tone failure might slightly annoy them. Weight accordingly.
• Gating vs scoring: safety/correctness dimensions should be conjunctive gates (must pass), not weighted scores (can be compensated by other dimensions). See the gate pattern below — do not fold a hard requirement (PII, safety, legal) into a WeightedSumAggregator, because a high score elsewhere can mask the violation.

Gates vs weighted scores (conjunctive requirements)

A WeightedSumAggregator lets dimensions compensate each other: a perfect JSON score can drag a PII leak up to "passing." For any requirement that must never be traded off (PII, safety, legal compliance), implement a gate — if it fails, the whole sample fails regardless of the other scores. OpenJudge ships only WeightedSumAggregator, so write a tiny gate aggregator:

from typing import Dict
from openjudge.runner.aggregator.base_aggregator import BaseAggregator
from openjudge.runner.aggregator.weighted_sum_aggregator import WeightedSumAggregator
from openjudge.graders.schema import GraderResult, GraderScore

class GatedWeightedSumAggregator(BaseAggregator):
    """Weighted sum that hard-fails (score=0.0) if any gate grader is below threshold.

    gate_graders: grader names that act as conjunctive gates (must pass).
    A gate 'passes' when its score >= gate_threshold.
    """

    def __init__(self, name: str, weights: Dict[str, float],
                 gate_graders: list[str], gate_threshold: float = 1.0):
        super().__init__(name)
        self.gate_graders = gate_graders
        self.gate_threshold = gate_threshold
        # Score only the non-gate dimensions; gates are pass/fail, not weighted.
        self._scorer = WeightedSumAggregator(
            name=name,
            weights={k: v for k, v in weights.items() if k not in gate_graders},
        )

    def __call__(self, grader_results: Dict[str, GraderResult], **kwargs) -> GraderResult:
        for gate in self.gate_graders:
            res = grader_results.get(gate)
            if isinstance(res, GraderScore) and res.score < self.gate_threshold:
                return GraderScore(
                    name=self.name, score=0.0,
                    reason=f"GATE FAILED: {gate}={res.score} (< {self.gate_threshold}). "
                           f"Hard requirement violated; weighted score suppressed.",
                    metadata={"gate_failed": gate},
                )
        # All gates passed → weighted sum of the remaining (compensable) dimensions.
        return self._scorer(grader_results, **kwargs)

# Example: PII leakage is a gate; JSON-field correctness is the compensable score.
runner = GradingRunner(
    grader_configs={
        "fields_present": fields_grader,   # claim_id/status/amount present (deterministic)
        "no_pii": no_pii_grader,           # GATE: 1.0 = clean, 0.0 = PII present
    },
    aggregators=GatedWeightedSumAggregator(
        name="overall",
        weights={"fields_present": 1.0},   # only non-gate dims are weighted
        gate_graders=["no_pii"],           # PII leak ⇒ overall 0.0 no matter what
    ),
    max_concurrency=8,
)

Rule of thumb: if a stakeholder would say "I don't care how good the rest is, this can never ship if X happens," then X is a gate, not a weight.

Common Mistakes

• All LLM judges, no deterministic checks. Every LLM call adds cost and noise. ~30-50% of evaluation dimensions can be checked with code. Check those first.
• One judge evaluating 3+ things. A single holistic judge produces unactionable verdicts. "The response scored 3/5" tells you nothing about what to fix. Split into one judge per dimension.
• Likert scales. "Rate helpfulness 1-5" produces scores that look scientific but can't be calibrated — annotators disagree on 3 vs 4 far more than pass vs fail.
• No few-shot examples. Without examples, the judge model guesses what "pass" means in your context. The borderline example is the most important one.
• Judge uses the same model as the target. Self-evaluation bias inflates scores. Always use a different model (or at minimum a different model version).
• Equal weights for composite scores. Equal weights = "I don't know what matters." Derive weights from failure prevalence or business impact.

Next Skills

After 02-metric-design:

• 03-align-human: You have graders. Now calibrate TPR/TNR against human labels to know if the automatic evaluation is trustworthy.
• 04-eval-report: Run evaluation at scale and generate analysis with OpenJudge's DistributionAnalyzer.