CORE GUIDE
Evaluation environments and verifier design
An AI evaluation is only as credible as the environment, task contract and verifier that determine what counts as success.
Mental model
Treat an evaluation as an executable claim about product quality: a controlled task runs in a representative environment, evidence is captured, and one or more verifiers decide whether the required outcome actually occurred.
Why it matters
AI systems can produce persuasive transcripts while failing the underlying task. Agent evaluations are especially vulnerable because tools, state changes and multiple turns create many ways to appear successful without satisfying the real postcondition. A useful eval therefore measures the environment outcome, preserves the trajectory for diagnosis, and combines graders whose strengths match the property being tested.
01
Build the eval around observable success
Start with a task whose initial state and success criteria are explicit. Run the model plus its harness against an environment that exposes realistic tools and side effects. Record the trial, then grade the parts that matter: deterministic checks for machine-verifiable invariants, model-based graders for bounded qualitative judgments, and human review where policy or ambiguity remains material. Repeat trials when stochastic variation can change the result.
02
Example: a coding agent says the fix is done
A coding agent edits a repository and reports success. The verifier does not grade the sentence; it inspects the resulting files, runs targeted tests and checks that prohibited files were not changed. The transcript is still retained because it helps explain a failure, but the environment state and executable assertions decide whether the task passed.
Common failure modes
- Grading only the final natural-language answer instead of the environment outcome.
- Using one brittle grader for properties that need multiple independent checks.
- Running a single stochastic trial and treating it as a stable quality estimate.
Engineering heuristics
- Define success criteria before writing the evaluation prompt.
- Prefer executable or state-based checks for claims that can be measured directly.
- Keep trajectories for diagnosis, but separate diagnostic evidence from the final pass/fail contract.
Takeaways
- 01An eval is a task, environment, trial and verifier system—not just a prompt.
- 02Outcome checks are stronger than self-reported success.
- 03Multiple trials and complementary graders make evaluation evidence more trustworthy.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.