Skip to main content

Validation Checklists

Statistical, categorical, correlation, temporal, consistency, and privacy checks run in Phase 5.

Phase 5: Validation Checks​

Statistical Accuracy Checks​

For each profiled numeric field, compute and compare:

MetricProfiledGeneratedVarianceStatus
Sales avg$228.23[value][x]%✓/✗
Sales min$0.44[value]—✓/✗
Sales max$22,638[value]—✓/✗
Profit avg$28.67[value][x]%✓/✗
Quantity avg3.79[value][x]%✓/✗

Threshold: If any average deviates by more than 10% from profiled AND calibration escape hatch was not already triggered (user accepted variance), flag it for correction.

Injected rows: Rows modified by anomaly injection (Phase 4B) are excluded from the mean, seasonal and trend checks; report them separately (count and what changed). Imperfections (Phase 4C) are applied after this validation passes; afterwards re-run only the privacy, ID/key and cross-field checks.

Categorical Distribution Checks​

For each profiled categorical field, verify generated frequencies are within ±3 percentage points of profiled:

CategoryProfiled %Generated %Status
Office Supplies60%[x]%✓/✗
Furniture22%[x]%✓/✗
Technology18%[x]%✓/✗

Correlation Direction Checks​

For each profiled correlation pair, verify the generated data preserves the relationship:

PairProfiled DirectionGenerated DirectionStatus
Sales ↔ ProfitPositive[result]✓/✗
Discount ↔ ProfitNegative[result]✓/✗

Method: Sort generated data by field_A, split into quartiles, check that field_B's average moves in the expected direction. If violated, flag for regeneration of the dependent field.

Temporal Pattern Checks (if applied)​

  • Monthly averages follow the profiled seasonal curve (within ±15% of expected seasonal index)
  • Overall trend direction matches profiled direction (growth/decline/flat)
  • Day-of-week distribution matches profiled weights (within ±5pp)

Cross-Field Consistency Checks​

  • Every Order ID year matches its Order Date year
  • Every Ship Date ≥ its Order Date
  • Every City is valid for its State (no city/state mismatches)
  • Every State is valid for its Region (matches profiled mapping)
  • Every State is valid for its Country
  • Every calculated field is computed correctly from base fields (spot-check 5 random rows)
  • No orphaned foreign keys (every Returns.Order_ID exists in Orders)
  • No duplicate values in fields marked unique
  • Null rates match profiled rates within ±2 percentage points

Privacy Checks​

  • No real values copied from PII-like columns (name, email, phone, address, ID-like)
  • No real full name + real address + real phone/email combination
  • Synthetic names only (no celebrities, public figures)
  • Randomized street numbers if addresses present