Skip to main content

Eval Engineering Playbooks

Playbooks · Eval engineering overview · Golden datasets →

Implementation guides for the Eval Blueprint. G.A.I.N Evaluation covers principles. These playbooks are the how: data sources, scorers, and plane-specific gates.

THE CLAIM

Eval is a control system, not a scoreboard. Build golden cases first, score every plane that can fail, and promote production signals back into the dataset. One end-of-pipeline number is not enough.

Two playbook groups

GroupOverviewWhat you build
MethodsGolden datasets →Datasets, synthetic coverage, online/dynamic scoring, human review, LLM-as-judge
PlanesInput →Per-plane gates from Input ① through Outcome ⑧

Plus Further reading (external) for third-party eval tools and guides mapped to this series.


  1. Golden datasets: versioned, plane-tagged cases before automation
  2. Synthetic generation and Human review: expand coverage and calibrate rubrics
  3. LLM-as-judge: scale scoring once humans set ground truth
  4. Plane playbooks: gate each plane your system actually owns (start with failure modes you already see)
  5. Online & dynamic: score live traffic and feed incidents back into goldens

Bridge reading: Eval Blueprint · G.A.I.N Evaluation.

Method playbooks at a glance

PlaybookOne-line purpose
Golden datasetsVersioned, plane-tagged, risk-tiered case libraries
Synthetic generationEdge and adversarial coverage without polluting goldens
Online & dynamicLive sampling, shadow scoring, drift, promote-back loops
Human reviewSampling, rubrics, adjudication that anchors automation
LLM-as-judgeScaled judging with bias controls and human calibration

Plane playbooks at a glance

#PlaybookOne-line purpose
InputParsing, intent, injection resistance, PII before inference
DataSource freshness, lineage, access boundaries
ContextRetrieval precision, ranking, packing, abstention
ReasoningFaithfulness, conclusion quality, multi-step logic
ToolSelection, arguments, idempotency, schema compliance
MemorySession scope, TTL, cross-session leakage
ActionPolicy enforcement, authorization, audit before side effects
OutcomeTask success, clarity, usefulness, trust in the final response

Who should read what

RoleStart withThen
AI platformGolden datasets, LLM-as-judgeOnline & dynamic, relevant plane playbooks
Domain squadGolden datasets, Human reviewPlane playbooks for your surface (e.g. Context, Tool)
Governance / riskHuman review, ActionInput, Outcome
RAG / knowledgeData, ContextEval Blueprint
Agent / PGAR teamsTool, ActionPGAR Runtime, Memory

Golden datasets →