ML Experiment Evaluation

Use this skill to choose how to evaluate machine-learning product changes before they consume live experiment traffic or affect users. It focuses on offline evaluation, offline-online correlation, interleaving, model filtering, and when classic A/B testing or adaptive strategies are justified.

Source Traceability

Primary source: Next-Level A/B Testing by Leemay Nassery. Guidance is transformed and paraphrased from Chapter 4 on offline evaluation, offline-online correlation, multi-armed bandits, and interleaving for rankers.

Related skills:

• experiment-sensitivity-optimization for reducing live variants and traffic.
• adaptive-experimentation-strategy for bandits and dynamic allocation.
• ab-test-design-brief for standard online A/B test planning.

Reference Routing

Need	Read
ML evaluation concepts	references/core/knowledge.md
Selection and validation rules	references/core/rules.md
Evaluation strategy examples	references/core/examples.md
Step-by-step evaluation plan	workflows/choose-ml-evaluation-strategy.md

Workflow

1. State the model change and product decision.
2. Identify the user harm or trust risk if a poor model reaches production.
3. Choose the lowest-cost evaluation that can filter bad candidates.
4. Check offline metrics and whether they correlate with online outcomes.
5. Use interleaving when ranker comparison needs high sensitivity with fewer users.
6. Escalate to online A/B testing or adaptive testing only when live evidence is needed and infrastructure can support it.

Output Format

# ML Evaluation Strategy

## Model Decision
[What model or ranking decision must be made.]

## Recommended Evaluation Path
[Offline only | Offline then A/B | Interleaving | A/B test | Adaptive strategy]

## Why
- Product risk:
- Offline signal available:
- Online evidence needed:
- Traffic or capacity constraint:

## Metrics
| Metric | Offline/Online | Role | Concern |
|--------|----------------|------|---------|

## Implementation Notes
- Data needed:
- Logging needed:
- Correlation check:
- Rollout guardrails:

Quality Bar

• Do not send poor offline candidates to live users just to get online evidence.
• Do not trust offline metrics until their relationship to online outcomes is understood.
• Do not use interleaving unless the product has a ranking or choice context where attribution can be logged.
• Do not recommend adaptive methods without checking data freshness, observability, and operational ownership.