Evaluation Reality Check

Evaluation Reality Check

Mission

Design evaluation strategies that match task type, deployment setting, time, groups, and decision risk.

Current pain point: Offline evaluation often overstates real-world performance because splits, metrics, and benchmarks do not match deployment.

Why this skill exists: Evaluation must simulate the future decision surface, not just produce a leaderboard score.

Operating Rules

• Start by restating the decision this skill is supporting.
• Inspect local artifacts first before asking for context.
• Treat scripts as evidence collectors, not as substitutes for judgment.
• Preserve raw data, notebooks, configs, and model artifacts unless the user explicitly asks for mutation.
• Mark missing context as unknown, not provided, or owner decision; do not invent it.
• Classify each blocker as stop, fix-first, monitor, accepted risk, or owner decision.
• Use the output contract exactly unless the user asks for a different format.
• Keep final recommendations auditable: every important claim needs evidence, assumption, or caveat.

What This Skill Must Do

• Transform a vague request into an explicit Evaluation Reality Check decision with named owner, evidence, assumptions, and action threshold.
• Inspect local artifacts first: data extracts, schemas, notebooks, SQL, pipeline configs, model reports, tickets, and prior decisions.
• Separate mechanical checks from expert judgment so another reviewer can reproduce what was checked and what was inferred.
• Classify findings as stop, fix-first, monitor, accepted risk, or owner decision instead of producing generic advice.
• Preserve raw evidence and never silently mutate data, notebooks, model artifacts, or production configs.
• Return a decision artifact that can be handed to a data scientist, ML engineer, governance reviewer, or stakeholder without hidden context.

Required Inputs

• task type
• data generation process
• deployment use
• target and features
• metric
• risk level

If an input is missing, inspect available files first. Ask only for information that cannot be recovered from the workspace and would change the recommendation.

Workflow

1. Identify whether the problem is tabular, time series, NLP, CV, ranking, recommendation, or anomaly detection.
2. Choose split strategy: temporal, grouped, stratified, rolling, external validation, or benchmark holdout.
3. Check leakage, label delay, distribution shift, uncertainty, and subgroup coverage.
4. Define confidence intervals, calibration, threshold testing, and failure slices.
5. Return an evaluation protocol and minimum launch evidence.

Red Flags

• random split for temporal deployment
• test set reused for tuning
• no external validation for domain shift
• single aggregate metric
• label delay ignored

When a red flag appears, slow down and surface it in Risks. A red flag does not always mean stop, but it must change the recommendation or the confidence level.

Resources

• Read references/playbook.md for the skill-specific checklist, scoring rubric, and failure modes.
• Read references/acceptance-tests.md before forward-testing clean, messy, and adversarial requests.
• Read references/agent-portability.md when adapting this skill to Claude, Gemini, Copilot, Cursor, Windsurf, Gravity, LangGraph, CrewAI, AutoGen, or local agents.
• Read references/quality-rubric.md when reviewing whether the output meets senior data-science standards.
• Use scripts/quick_validate_skill.py . --strict before publishing or installing the skill.

Use only the specific reference file needed for the task; keep context small.

Output Contract

Return these sections unless the user requests another format:

• Problem: the decision, artifact, model, data source, or workflow being handled.
• Inputs: files, data, stakeholder context, assumptions, and missing context used.
• Checks performed: concrete inspections, scripts, and reasoning checks.
• Findings: prioritized observations with evidence.
• Decision: go, fix-first, stop, proceed-with-risk, or needs-owner-review.
• Risks: unresolved blockers, caveats, and owner decisions.
• Next actions: smallest useful follow-up steps.
• Artifacts: generated files, specs, reports, scripts, or links.

Final Checks

• Did the response answer the actual decision, not just analyze data?
• Did it preserve evidence and avoid silent mutation?
• Did it name unknowns and owner decisions?
• Did it include at least one concrete next action?
• Did it avoid overclaiming causality, readiness, compliance, or production safety?