| name | golden-dataset-evaluation |
| description | Use this skill when changing prompts, tools, models, or agent orchestration so the resulting behavior is evaluated against the synthetic preferred-response golden dataset before sign-off. |
| license | MIT |
| metadata | {"author":"perficient_interview","inspired_by":"addyosmani/agent-skills test-driven-development","version":"1.0"} |
Golden Dataset Evaluation
Overview
This skill treats preferred responses as the proof layer for agent quality. Agent changes are incomplete until they have been evaluated against the curated synthetic goldens.
Instructions
- Load the golden dataset from
data/synthetic_preferred_responses.json.
- For each case, compare the agent response against:
- The preferred response
- The required references in
must_include
- The expected issue and policy IDs
- Score on four axes:
- factual alignment
- policy grounding
- actionability
- required-reference coverage
- Fail any case that omits a critical policy or cites the wrong issue or escalation path.
- Save machine-readable results for later regression checks.
Production Guardrail
Synthetic goldens are acceptable for prototype evaluation.
For production, historical operator responses must be cleaned, deduplicated, and policy-validated before they become judge references.