Llm As Judge

---

name: llm-judge-ensemble version: 1.0.0 description: Build a cost-efficient LLM evaluation ensemble with sampling, tiebreakers, and deterministic validators. Learned from 600+ production runs judging local Ollama models. homepage: https://github.com/reddinft/skill-llm-as-judge metadata: { "openclaw": { "emoji": "⚖️", "requires": { "bins": ["python3", "ollama"], "env": ["ANTHROPIC_API_KEY", "OPENAI_API_KEY"] }, "primaryEnv": "ANTHROPIC_API_KEY", "network": { "outbound": true, "reason": "Calls Anthropic (Claude Sonnet) and OpenAI (GPT-4o-mini) as judge models at 15% sampling rate. Optional Gemini tiebreaker. Local model inference via Ollama stays on-device." } } }

---

LLM-as-Judge

Build a cost-efficient LLM evaluation ensemble for comparing and scoring generative AI outputs at scale.

When to Use

Evaluating generative AI outputs across multiple models at scale (100+ runs)

Comparing local/OSS models against cloud baselines in shadow-testing pipelines

Building promotion gates where models must prove quality before serving production traffic

Any scenario where deterministic tests alone can't capture output quality

When NOT to Use

One-off evaluations (just read the output yourself)

Tasks with deterministic correct answers (use exact-match or unit tests)

When you can't afford any external API calls (this pattern uses Claude/GPT as judges)

Architecture: Three-Layer Evaluation

Layer 1: Deterministic Validators (Free, Instant)

Run on 100% of outputs. Zero cost. Catches obvious failures before burning judge tokens.

JSON schema validation — does the output parse? Does it match the expected schema?

Regex checks — required fields present, format constraints met

Length bounds — output within acceptable min/max character count

Entity presence — do required entities from the input appear in the output?

If Layer 1 fails, score is 0.0 — no need to invoke expensive judges.

Layer 2: Heuristic Drift Detection (Cheap, Fast)

Run on 100% of outputs that pass Layer 1. Minimal cost (local computation only).

Entity overlap — what fraction of entities in the ground truth appear in the candidate?

Numerical consistency — do numbers in the output match source data?

Novel fact detection — does the output introduce facts not present in the input/context? Novel facts suggest hallucination.

Structural similarity — does the output follow the same structural pattern as ground truth?

Layer 2 produces heuristic scores (0.0–1.0) that contribute to the final weighted score.

Layer 3: LLM Judges (Expensive, High Quality)

Sampled at 15% of runs to control cost. Forced to 100% during promotion gates.

Two independent judges (e.g., Claude + GPT-4o) score the output. Each judge evaluates all 6 dimensions independently.

Tiebreaker pattern: When primary judges disagree by Δ ≥ 0.20 on any dimension, a third judge is invoked. The tiebreaker score replaces the outlier. This reduced score variance by 34% at only 8% additional cost.

The 6 Scoring Dimensions

| Dimension | Weight | What It Measures | |---|---|---| | Structural accuracy | 0.20 | Format compliance, schema adherence | | Semantic similarity | 0.25 | Meaning preservation vs ground truth | | Factual accuracy | 0.25 | Correctness of facts, numbers, entities | | Task completion | 0.15 | Does it actually answer the question? | | Tool use correctness | 0.05 | Valid tool calls (when applicable) | | Latency | 0.10 | Response time within acceptable bounds |

Weights are configurable per task type. Tool use weight is redistributed when not applicable.

Critical Lesson: None ≠ 0.0

When a dimension is not sampled (LLM judge not invoked on this run), record the score as null, not 0.0. Unsampled dimensions must be excluded from the weighted average, not treated as failures.

Early bug: recording unsampled dimensions as 0.0 created a systematic 0.03–0.08 downward bias across all models. The fix: null means "not measured", which is fundamentally different from "scored zero".

# WRONG — penalises unsampled dimensions
weighted = sum(s * w for s, w in zip(scores, weights)) / sum(weights)
RIGHT — exclude null dimensions
pairs = [(s, w) for s, w in zip(scores, weights) if s is not None]
weighted = sum(s * w for s, w in pairs) / sum(w for _, w in pairs)

Cost Estimate

With 15% LLM sampling, average cost per evaluated run: ~$0.003

Layer 1 + Layer 2: $0.00 (local computation)

Layer 3 (15% of runs): ~$0.02 per judged run × 0.15 = ~$0.003

Tiebreaker (fires ~12% of judged runs): adds ~$0.0003

At 200 runs for promotion: total judge cost ≈ $0.60 per model per task type.

Worked Example: Summarisation Evaluation

from evaluation import JudgeEnsemble, DeterministicValidator, HeuristicScorer
Layer 1: must be valid text, 50-500 chars
validator = DeterministicValidator(
    min_length=50,
    max_length=500,
    required_format="text",
)
Layer 2: check entity overlap with source
heuristic = HeuristicScorer(
    check_entity_overlap=True,
    check_novel_facts=True,
    check_numerical_consistency=True,
)
Layer 3: LLM judges (sampled)
ensemble = JudgeEnsemble(
    judges=["claude-sonnet-4-20250514", "gpt-4o"],
    tiebreaker="claude-sonnet-4-20250514",
    sample_rate=0.15,
    tiebreaker_threshold=0.20,
    dimensions=["structural", "semantic", "factual", "completion", "latency"],
)
Evaluate
result = ensemble.evaluate(
    task_type="summarize",
    ground_truth=gt_response,
    candidate=candidate_response,
    source_text=original_text,
    validator=validator,
    heuristic=heuristic,
)
print(f"Weighted score: {result.weighted_score:.3f}")
print(f"Dimensions: {result.scores}")  # {semantic: 0.95, factual: 0.88, ...}
None values for unsampled dimensions

Tips

Start with Layer 1 — you'd be surprised how many outputs fail basic validation

Log everything — store raw judge responses for debugging score disputes

Calibrate on 50 runs — before trusting the ensemble, manually review 50 outputs against judge scores

Watch for judge drift — LLM judges can be inconsistent across API versions; pin model versions

Force judges at gates — 15% sampling is fine for monitoring, but promotion decisions need 100% coverage on the final batch

Start with Layer 1 — you'd be surprised how many outputs fail basic validation

Log everything — store raw judge responses for debugging score disputes

Calibrate on 50 runs — before trusting the ensemble, manually review 50 outputs against judge scores

Watch for judge drift — LLM judges can be inconsistent across API versions; pin model versions

Force judges at gates — 15% sampling is fine for monitoring, but promotion decisions need 100% coverage on the final batch