What Should a Payroll Parallel-Run Variance Review Prove?

Research question: whether a payroll support team can identify and route parallel-run variances without deciding pay entitlement or authorizing payroll? This protocol tests a narrow administrative evidence model for Philippines staffing operations. It is not legal advice, an employment decision, a security certification, or a promise that the same workflow fits every organization.

Why the question matters: A buyer must distinguish differences caused by source timing, mapping, rounding, configuration, approved correction, or unexplained output. An agreeing total can conceal person-level errors, while a large difference can be expected and fully evidenced.

Evidence frame: the Philippine Data Privacy Act implementing rules describe transparency, legitimate purpose, proportionality, accountability, security, access, retention, and responsibilities around outsourced processing. DOLE advisories provide current Philippine labor context. NIST CSF 2.0 and SP 800-53 supply general governance, identity, access, audit, change, and risk-control language. They are inputs to a buyer’s design, not a substitute for facts, contracts, applicable law, or accountable professional judgment.

Unit of analysis: one worker, one pay component, one period, and two named outputs produced from versioned inputs. Fixing the unit before testing prevents a favorable batch total from hiding one unresolved person, instruction, record, or downstream handoff. Every case receives a stable fictional reference and every conclusion must point to an observable source event.

Test set: create one hundred and twenty fictional comparisons covering joiners, leavers, overtime, allowances, deductions, retroactive changes, duplicates, late approvals, rounding, mapping changes, missing records, and offsetting errors. Use invented people, organizations, amounts, accounts, documents, and identifiers only. A study administrator keeps the seeded answer key separate until both reviewers finish their first pass.

Minimum fields: worker token, period, component, source version, cutoff, input, baseline output, candidate output, variance, tolerance rule, evidence link, reviewer, owner, disposition, correction version, release decision, and acknowledgment. Define the purpose and allowed values for every field. Blank, unknown, not applicable, not yet received, restricted, and cannot determine remain distinct states. Reviewers may not turn absence into a convenient answer.

Before review, the accountable business owner freezes the population, source hierarchy, state definitions, permitted actions, access roles, response windows, serious-error classes, and stop conditions. A later policy change creates a new version and a targeted rerun; it never silently rewrites the original observation.

Primary measure: correct classification and routing of matched, expected difference, unexplained difference, stale source, duplicate, corrected, approved, or cannot determine states; report person-level errors separately from net totals. Reviewers must cite the exact evidence used for each state. A confident guess counts as an error even when it happens to match the seeded answer.

Error model: Net-total agreement is not person-level accuracy. A variance within tolerance is not automatically correct. A corrected input does not prove that the receiving system used it, and a coordinator cannot convert an unexplained difference into an approved adjustment.

Decision boundary: The coordinator may assemble versioned inputs, run declared comparisons, classify observable differences, preserve evidence, and route exceptions. Payroll, finance, HR, tax, legal, and authorized advisers decide entitlement, rules, adjustments, release, and communication.

Controlled comparison: Compare aggregate-only reconciliation with worker-and-component comparison linked to source versions; seed offsetting errors and late changes Give both workflows the same underlying cases in randomized order. Compare correctness, unnecessary access, unresolved work, serious errors, and review time rather than relying on completion speed alone.

Privacy and security treatment: Use invented workers and values. Restrict payroll detail, bank information, government identifiers, and exports. Test access, notifications, downloads, backups, and deletion. Record who can view, change, export, and delete each artifact. Test linked systems and notification paths because a restricted main record can still leak through email, calendars, downloads, integrations, or backups.

Include positive controls that should proceed, negative controls that should stop, and ambiguous controls that should reach an authorized owner. A workflow that never stops is not controlled; one that stops every case is not operationally useful. Keep the expected route and reason in the hidden answer key.

Analysis plan: Report variance counts and values by class, component, source version, and cutoff distance. Separate false match, false exception, stale input, duplicate, and unsupported closure. Two reviewers independently classify an overlapping sample. Preserve disagreements and resolve them through the named owner. Do not average classifications or let the first entry become authoritative merely because it appeared first.

Set acceptance thresholds before opening the answer key. Define the minimum routing accuracy, maximum unresolved age, maximum tolerated disclosure, and failures that stop the pilot. Report counts with denominators and list exclusions with reasons. Faster handling cannot compensate for an unauthorized decision, sensitive-data exposure, or false closure.

Run a repeatability check with a second reviewer who receives the written rules and clean cases but no coaching. Low agreement indicates unclear rules, missing evidence, or inconsistent source access. Version the clarification and rerun affected cases; do not label every disagreement as an individual training problem.

Add a temporal test after the static review. Replay selected cases when a cutoff passes, an approver changes, a source is corrected, or a downstream acknowledgment arrives late. The expected state should change only when the declared transition evidence exists. Record who observed the event, which rule version applied, and whether notifications or dependent systems updated. This catches designs that look accurate in a snapshot but cannot preserve history or distinguish an overdue item from a superseded one.

Assess operational recovery as well as normal processing. Remove one required source, delay one owner, introduce one duplicate, and make one integration temporarily unavailable. The coordinator should preserve the last known state, state what cannot be determined, avoid reconstructing missing facts from memory, and route the case through the approved contingency path. Measure whether work resumes from preserved evidence without double action, unauthorized disclosure, or silent closure when the source returns.

Add a component-level causality review before disposition. For every variance, reconstruct the sequence from frozen source input through mapping, calculation, rounding, approval, import, and candidate output. Test at least one offsetting pair where the batch total agrees but two workers are wrong, one late approval that belongs to a later cutoff, and one corrected source that never reaches the calculation engine. Require the payroll owner to distinguish expected configuration behavior from an authorized correction; the coordinator records that decision without inventing a tolerance. The final evidence should show whether the rerun used the corrected version, whether the original output remains preserved, and whether release approval refers to the exact compared files. This makes the study sensitive to false reconciliation rather than merely visible arithmetic differences.

Ask a prospective provider for artifacts that match the operating claim: a sanitized workflow demonstration, blank register, role-permission view, change history, exception map, and sample audit export. Each artifact has its own date and scope. Marketing statements, policy documents, and successful demonstrations are point-in-time evidence, not proof of continuous operation.

Separate observed fact, rule-based classification, accountable-owner decision, and researcher inference in the final table. Preserve missing events, integration delays, inaccessible sources, ambiguous definitions, and unavailable owners as explicit uncertainty. “Cannot determine” is a useful result when the source does not support a stronger statement.

Limitations: Synthetic calculations cannot establish lawful pay, tax, contribution, exchange-rate, or contractual treatment. Platform configuration and payroll calendars vary. Begin any live pilot with a small approved queue, named reviewers, least-privilege access, monitored exceptions, and a stop rule for unexpected sensitive data or work outside the written lane.

Decision-grade output: a parallel-run variance register connecting each difference to both outputs, source versions, classification evidence, owner disposition, correction lineage, and release decision. A buyer can use the artifact to compare operating discipline, but it does not guarantee outcomes or transfer accountability from the responsible organization.

Sources checked September 25, 2026: National Privacy Commission, Implementing Rules and Regulations of the Data Privacy Act of 2012 (https://privacy.gov.ph/implementing-rules-regulations-data-privacy-act-2012/); National Privacy Commission, The Data Privacy Act and Its IRR (https://privacy.gov.ph/the-data-privacy-act-and-its-irr/); Department of Labor and Employment Bureau of Working Conditions, Labor Advisories (https://bwc.dole.gov.ph/issuances/labor-advisories/); National Institute of Standards and Technology, Cybersecurity Framework 2.0 (https://www.nist.gov/cyberframework); National Institute of Standards and Technology, Security and Privacy Controls for Information Systems and Organizations, SP 800-53 Revision 5 (https://csrc.nist.gov/pubs/sp/800/53/r5/upd1/final). These primary government sources frame privacy, labor context, governance, access, audit, and risk questions. They do not decide a specific employment matter, certify a provider, or replace advice from authorized legal, HR, payroll, security, benefits, or finance owners.