ometere123/the-skills-bible/ai-evaluation-architect

AI Evaluation Architect

Purpose

Create trustworthy evidence for model selection, release, monitoring, and improvement rather than relying on demos or one benchmark.

Use this skill when

• An AI system needs an evaluation strategy.
• Several model, prompt, retrieval, or agent variants must be compared.
• Release thresholds and regressions need governance.
• Human and automated evaluation must be coordinated.

Do not use this skill as the primary workflow when

• The task is one narrow LLM evaluator.
• No intended task or acceptable outcome is defined.
• The user needs model training rather than evaluation architecture.

Required inputs

• AI system outcomes and risks.
• Models, prompts, tools, data, and workflows.
• User populations and operating conditions.
• Historical failures and production feedback.
• Release, cost, and governance requirements.

Classify missing information as blocking, important, or optional. Inspect the actual product requirements, workflows, data, models, prompts, tools, permissions, evaluations, traces, incidents, policies, and production constraints before asking for information that can be verified directly.

Core operating principles

• Start from the decision, workflow, user outcome, or automation need rather than from a model or vendor.
• Use the simplest deterministic method that can satisfy the requirement before adding probabilistic components.
• Separate model capability, application logic, tools, memory, retrieval, policy, and human authority.
• Treat prompts, datasets, tools, schemas, model versions, policies, and evaluations as versioned production artefacts.
• Ground consequential outputs in authoritative evidence and expose uncertainty, provenance, and limitations.
• Constrain permissions, data access, actions, spending, recursion, and runtime before granting autonomy.
• Evaluate representative tasks, edge cases, adversarial inputs, failures, and affected user groups before release.
• Preserve privacy, security, accessibility, human agency, appeal, and safe failure throughout the lifecycle.
• Make model and agent behaviour observable without logging secrets, private reasoning, or unnecessary personal data.
• Design rollback, fallback, escalation, incident response, and decommissioning before production deployment.

Workflow

1. Define evaluation decisions

• State whether evidence supports selection, release, routing, monitoring, incident response, or retirement.
• Avoid evaluation without a decision.

2. Build task taxonomy

• Group ordinary, difficult, rare, adversarial, safety, subgroup, temporal, and failure tasks.
• Map user impact.

3. Create datasets

• Use held-out, challenge, production-derived, synthetic, counterfactual, and human-authored examples with provenance and contamination controls.

4. Define oracles

• Choose exact labels, reference answers, evidence support, independent calculations, simulations, human rubrics, or outcome measures.

5. Choose metrics

• Use task quality, calibration, retrieval, tool success, safety, fairness, latency, cost, abstention, and human effort.
• Avoid one score.

6. Design human evaluation

• Define rubrics, rater qualification, blinding, disagreement, adjudication, and quality checks.
• Protect raters.

7. Design automated evaluation

• Use deterministic checks, model graders, simulations, and metamorphic tests with validation against humans.
• Avoid circular grading.

8. Set thresholds

• Define minimums, non-regression, slice limits, uncertainty, and release exceptions based on impact.

9. Integrate lifecycle

• Run development, pre-release, canary, post-release, periodic, and incident-triggered evaluations on exact versions.

10. Govern evidence

• Version suites, prevent leakage, publish evaluation cards, review stale tasks, and record residual gaps.

Decision rules

Required outputs

• Evaluation-decision map.
• Task taxonomy.
• Evaluation datasets.
• Oracle design.
• Metric portfolio.
• Human-evaluation protocol.
• Automated evaluation.
• Thresholds.
• Lifecycle integration.
• Evaluation governance.

AI system quality gates

• Problem gate: The intended outcome and the reason to use AI are explicit.
• Data gate: Sources, rights, quality, lineage, privacy, and retention are controlled.
• Model gate: Model capability, limits, version, configuration, and fallback are documented.
• Grounding gate: Claims and actions use suitable evidence, tools, or deterministic checks.
• Authority gate: The model cannot exceed the permissions, spending, scope, or actions granted by policy.
• Evaluation gate: Representative, adversarial, failure, and subgroup evaluations meet approved thresholds.
• Human gate: High-impact, ambiguous, irreversible, or exceptional outcomes receive appropriate human review.
• Safety gate: Abuse, injection, privacy, harmful content, and unsafe action paths are tested and bounded.
• Observability gate: Inputs, outputs, tools, state, latency, cost, and outcomes are diagnosable with data minimisation.
• Lifecycle gate: Versioning, rollout, rollback, monitoring, incident response, and retirement are operational.

Evidence standard

For consequential AI conclusions, identify the exact model, model version, system prompt, tool set, retrieval index, dataset, policy, evaluation set, configuration, environment, and time window. Distinguish observed behaviour, measured estimates, supported inference, and unknowns. Do not claim capability, safety, fairness, or reliability from a few demonstrations.

Failure handling and escalation

Pause or escalate when:

• The system's intended authority, affected users, or acceptable error is unclear.
• Training, retrieval, or evaluation data lacks provenance, permission, or privacy controls.
• A model or agent may send messages, move money, change records, execute code, or access sensitive systems beyond approved scope.
• Prompt injection, tool compromise, data poisoning, identity confusion, or uncontrolled recursion is plausible.
• A high-impact outcome cannot be independently verified or appealed.
• Evaluation evidence is missing, stale, non-representative, or materially regresses.
• Residual safety, privacy, legal, financial, or operational risk exceeds delegated authority.

Report the exact blocker, verified state, affected outcome, safe options, and smallest required decision.

Common anti-patterns

• One public benchmark as release evidence.
• Test set used during prompt iteration.
• Model grading itself without validation.
• Aggregate score hiding critical subgroup failure.
• No cost or latency measurement.
• Human raters given vague preference instructions.
• Evaluation suite never refreshed.

Completion checklist

• [ ] Every evaluation supports a decision.
• [ ] Tasks represent real and adverse conditions.
• [ ] Oracles are independent enough.
• [ ] Metrics include safety and operations.
• [ ] Thresholds are risk-based.
• [ ] Exact system versions are recorded.

AI implementation record

# AI Implementation Record
 
## Outcome
[Task, user need, decision, or automation result]
 
## Scope and authority
[Users, systems, data, tools, permissions, budgets, and exclusions]
 
## Architecture
[Models, prompts, retrieval, memory, tools, policies, humans, and fallback]
 
## Data and provenance
[Sources, rights, lineage, privacy, retention, and quality]
 
## Evaluation
[Tasks, datasets, metrics, thresholds, adversarial cases, and subgroup results]
 
## Safety and security
[Injection, abuse, harmful output, tool constraints, privacy, and escalation]
 
## Deployment
[Versioning, routing, rollout, rollback, compatibility, and cost]
 
## Observability
[Traces, outcomes, latency, tokens, tool calls, incidents, and redaction]
 
## Residual risk
[Known limits, owner, acceptance, expiry, and user communication]
 
## Follow-up
[Monitoring, retraining, index refresh, policy updates, and decommission triggers]

Related skills

• llm-evaluation-engineer
• agent-evaluation-engineer
• ai-red-team-engineer
• model-lifecycle-governance-engineer

Maintenance rules

Review this skill when models, datasets, policies, tools, regulations, or workflows change; evaluations or incidents reveal a missing safeguard; the workflow becomes tied to one vendor; the file approaches 500 lines; or reusable schemas, evaluators, test sets, or guardrail policies would improve execution.