Payroll review · Synthetic ERP data

Construction Payroll

Find likely pay errors before release. Review the calculation, approval status and exact timesheet records behind each flag.

Model accuracy — deterministic rule validation

100% precision and recall is expected here: these deterministic checks are validated against synthetic labels built from the same rules. This is not a statistical model, an independent benchmark or evidence of real-world predictive performance.

Rule-level results and label overlap

The answer key assigns one primary error type per row. Rules can overlap: every overtime miscalculation also fails the expected-pay check. The Wrong Rate rule therefore has 50% exact-type precision, even though all its flags are genuine pay errors.

Rule	Flagged	Primary labels	Matching labels	Exact-type precision	Exact-type recall

Pay-error worklist

Ranked by number of triggered rules, then absolute expected-pay difference, then gross pay. This is review priority, not a probability or a savings estimate.

Find employee or recordError type

Employee / record	Role / site	Pay period	Regular / OT hours	Gross pay (INR)	Reasons for review

How the checks work

Regular hours exclude overtime. Expected gross pay = regular hours × contracted hourly rate + overtime hours × 1.5 × contracted hourly rate. A difference of up to ₹0.50 is accepted because the source rounds gross pay to whole rupees.

Duplicate Hours: more than one record for the same employee and pay-period end date. Both records are flagged for review; this alone does not prove duplicate payment.
Wrong Rate: gross pay differs from expected pay by more than ₹0.50. This flags a calculation mismatch; it does not establish the underlying cause.
Missing Approval: approval status is anything other than Approved.
Overtime Miscalculation: overtime hours are positive but gross pay is at or below straight-time pay for all hours, within ₹0.50 tolerance. No overtime premium is evident.

The corrected workbook contains 40 duplicate pairs (80 records) and two corrected zero-overtime seed cases. Answer-key labels are used only for validation, never to generate flags. Review findings before changing or releasing pay.