tikalk/adlc-team-skills/evals-validate

evals-validate

What this skill does

Conducts comprehensive validation of the implemented evaluation system following EDD principles to ensure production readiness through statistical analysis, performance verification, and quality assurance.

Output:

1. Statistical Validation - TPR/TNR analysis, accuracy metrics, confidence intervals
2. Performance Validation - SLA compliance verification for evaluation pyramid tiers
3. Quality Assurance - Goldset integrity, example balance, coverage analysis
4. Holdout Dataset Validation - Unbiased accuracy assessment on reserved test set
5. Auto-handoff to /evals-analyze for closed loop trajectory analysis

Key EDD Principles Applied:

• Principle IV: Evaluation Pyramid - Tier performance SLA validation (Tier 1 <30s, Tier 2 <5min)
• Principle II: Binary Pass/Fail - Statistical compliance verification
• Principle IX: Test Data as Code - Holdout dataset validation integrity
• Principle III: Error Analysis - Pattern stability validation

When to use

• After `/evals-implement`: Execute the evaluation suite and measure quality
• CI/CD Pipeline gate: Run evaluations before release to ensure no regressions
• Periodic audit: Verify evaluator accuracy on holdout data to check for model drift

When NOT to use

• Evaluator not generated: Run /evals-implement to build grader files first
• Analysing failure traces: Use /evals-analyze to extract deep insights from run results

Process

User Input

$ARGUMENTS

• --holdout-only — Validate only on holdout dataset (unbiased validation)
• --performance-only — Skip statistical analysis, focus on SLA compliance
• --metrics METRICS — Specific metrics to validate (tpr, tnr, accuracy, performance)

Execution Steps

Phase 1: Execute Evaluations

Runs the underlying framework CLI directly:

• PromptFoo: npx promptfoo eval --config evals/promptfoo/config.js
• DeepEval: pytest evals/deepeval/ -v or python evals/deepeval/config.py

Phase 2: Compute Statistical Validation

• Parse generated results JSON from evals/results/.
• Calculate True Positive Rate (TPR) and True Negative Rate (TNR).
• Calculate overall accuracy with 95% confidence intervals.
• Ensure no Likert scales or numerical scores leak into results.

Phase 3: SLA Compliance Check

• Measure execution times for Tier 1 and Tier 2.
• Verify Tier 1 completes under 30 seconds.
• Verify Tier 2 completes under 5 minutes.
• Check headroom analysis (SLA budget consumed).

Phase 4: Write Validation Report

• Write validation results to evals/results/validation_report.md.
• Include pass/fail counts, TPR/TNR table, SLA timings, and holdout set results.

Phase 5: Auto-Handoff

Trigger /evals-analyze to close the loop.

Verification

• Evaluation execution successfully completed with results JSON written to evals/results/
• evals/results/validation_report.md created with TPR/TNR and SLA metrics
• Statistical metrics calculated with confidence intervals
• Headroom and SLA compliance verified
• Handover summary lists results and validation report path