LLMEvaluationMethodology
title: LLM Evaluation Methodology radar_quadrant: Techniques radar_ring: Assess radar_position: inner
LLM Evaluation Methodology is the practice of defining measurable success criteria and building structured evaluation suites before deploying LLM-based applications. The approach is documented in Anthropic's official guidance ("Define success criteria and build evaluations") and is central to the prompt engineering cycle: define criteria, build evals, measure, iterate.
Good success criteria share four properties. They are specific — "accurate sentiment classification" rather than "good performance." They are measurable — quantitative metrics or consistently-applied qualitative scales, including for subjective topics: safety criteria can be expressed as "fewer than 0.1% of outputs flagged for toxicity across 10,000 trials." They are achievable — grounded in industry benchmarks or prior experiments rather than aspirational targets beyond frontier model capability. They are relevant — aligned with real user needs; citation accuracy is critical for medical applications but less so for casual chatbots.
Most production applications require multidimensional evaluation across several criteria simultaneously: task fidelity, consistency, relevance and coherence, tone and style, privacy preservation, context utilisation, latency, and cost.
Be task-specific. Evals must mirror the real-world task distribution, including edge cases. Generic benchmarks do not predict performance on a specific deployment's inputs.
Automate when possible. Structure questions to enable automated grading: multiple-choice, string match, code-executed assertions, or LLM-graded responses. Automated grading scales; human grading does not.
Prioritise volume over quality. More automated evals at moderate signal strength outperforms fewer hand-graded evals. A 500-question automated suite surfaces regressions faster than a 20-question expert-reviewed set.
Use LLM-as-judge for subjective criteria. For tone, coherence, or safety — criteria that resist string matching — a grading model evaluating responses against a rubric provides scalable subjective assessment. This pattern is covered in more detail in the Year of Building with LLMs blip.
Placed in Techniques / Assess / inner.
Eval methodology is the gap most LLM projects skip in the early stages and regret later. Without defined success criteria before building, prompt changes cannot be evaluated objectively — teams ship on vibes and discover regressions in production. The Anthropic framework is provider-agnostic; it applies equally to projects using Claude, OpenAI, or local models via Ollama.
This technique is distinct from inspect_ai (Languages & Frameworks / Assess / inner), which is the framework tool implementing evals. LLM Evaluation Methodology is the upstream design practice — what to measure and how to structure the measurement — that inspect_ai (or any eval harness) then implements.
Directly complements: inspect_ai (eval framework), Year of Building with LLMs (LLM-as-judge pattern), RAG Chunking Strategies (chunking evals), and DuckDB Vector Search (retrieval precision measurement).
Trial gate: success criteria defined and an automated eval suite built before or alongside the first prompt change in a real LLM application, with at least one regression caught by the evals before it reached production.
