Automated screening change logs for recruiting assistants

Research question: How can a recruiting assistant preserve evidence needed to review changes in an automated candidate-screening workflow?

A screenshot of today’s setting cannot establish which rule processed yesterday’s candidate. When a threshold, question, model or default changes without an effective-time boundary, later review cannot reliably reconstruct the population or decision path.

This report studies a bounded work lane for a Philippines-based recruiting assistant. It does not grade a worker, provider, profession, country or software product. The question is whether a buyer can define a traceable administrative process while keeping consequential judgment with the correct owner.

The observation unit is one screening configuration period linked to job, criteria, tool version, effective time, input fields, threshold, population, accommodation route, human-review rule, approver, notices, outcomes, exceptions and superseding configuration. A fixed unit makes omissions inspectable and prevents a completion label from hiding an unresolved decision.

Employment guidance and the limits of configuration evidence

EEOC materials identify technology-assisted recruiting as an enforcement concern and describe disability risks from algorithmic tools.^[7][8] The Uniform Guidelines discuss selection-procedure evidence and records.^[9] These sources do not approve a tool or authorize an assistant to validate it.

The direct sources are identified by publisher, title, URL and checked date.^[7][8][9] Authority matters more than source count: a page that repeats another source is not independent evidence. Publication and access dates also mean different things; the checked date records this review, not the effective date of every underlying requirement.

The EEOC sources identify risks and recordkeeping concerns; they do not turn a change log into a validation study. An employer must connect this design to the actual job analysis, selection procedure, accommodation process, retention duties, vendor contract and applicable jurisdiction before candidates are processed.

Separate populations at the configuration boundary

Record the requisition, job-analysis reference, approved criteria, tool and vendor-supplied version, threshold, effective timestamp and change approver. Do not claim visibility into undocumented vendor internals.

Map each candidate to the configuration period that processed the record. Preserve reruns, errors, overrides and the difference between an automated recommendation and the final human decision.

Reconcile the populations entering the stage, receiving a score, failing technically, requesting accommodation, withdrawing, receiving human review, advancing and receiving disposition. Keep restricted details out of general reports.

Compare periods only after considering job, recruiting channel, applicant mix, missingness, timing and accommodation handling. Route job relatedness, validation, adverse impact and candidate decisions to qualified owners.

Join candidate history to tool-version history

Create a structured record around one screening configuration period linked to job, criteria, tool version, effective time, input fields, threshold, population, accommodation route, human-review rule, approver, notices, outcomes, exceptions and superseding configuration. Use controlled statuses, named owners and stable identifiers. “Done” should mean the defined destination state was checked, not merely that a message, upload or request was sent.

Configuration evidence must be versioned rather than overwritten. Preserve the prior threshold, questions, model or rule set with its effective interval; add the approved successor; and map every processed candidate to the version actually used. Restricted applicant details belong in controlled systems, with the audit log carrying stable references and aggregate states.

The reader gets a version-control and denominator design that makes silent configuration drift inspectable while keeping legal, statistical and employment judgment outside the assistant lane.

Analyze a screening change without inventing causation

Choose the population boundary before reading outcomes. For each applicant, identify the requisition, stage-entry time, tool version, threshold, completion state, technical interruption, requested accommodation path, human override and final stage state. Reconcile entrants to exits so withdrawals and missing dispositions cannot silently leave the denominator.

A before-and-after difference is descriptive, not an explanation. Applicant mix, recruiting source, job requirements, labor-market conditions, recruiter behavior, test completion and another simultaneous configuration change may differ between periods. Report those changes beside the counts and ask the qualified owner whether a different design or further analysis is required.

Configuration evidence also has layers. A vendor release date, tenant enablement time, administrator save event and approval can be four different moments. The change log should preserve all four when available. If the effective state for a candidate cannot be reconstructed, classify that record as unknown rather than assigning it to the cleaner comparison group.

Threshold-change case: stop a mixed candidate population

A vendor changes a default ranking threshold while applications remain open. The release notice lacks the tenant’s effective time, and the administrator log shows a save event two days later. The assistant freezes both records, identifies the last known candidate under the earlier state and the first known candidate under the later state, and marks the interval between them unresolved.

Candidates are separated by observed configuration rather than forced into calendar-week groups. Technical failures, withdrawals, requests routed through the restricted accommodation process and human overrides remain visible. The assistant pauses an unapproved bulk disposition because the hiring team has not decided how to handle people who may have encountered different rules.

Employment and accessibility owners decide whether review, reprocessing or another action is appropriate. The change log records their instruction and the affected population. It does not store disability details in the general project sheet or describe a stage-rate difference as discrimination, fairness or validity. Those conclusions require qualified analysis beyond an administrative history.

To test the join, take an authorized sample from each configuration state and reconstruct the path without looking at the final hiring outcome first. Verify stage-entry time, assessment completion, score availability, threshold applied, human action and disposition. A record that cannot be joined belongs in an unknown bucket; silently dropping it can change the denominator.

Restricted information needs a different governance path from ordinary configuration data. The project register may show that an accommodation route was invoked and whether the candidate returned to the selection flow, but the request and supporting information stay with authorized personnel.

When owners approve remediation, preserve its scope. Reprocessing all candidates, reviewing only the unresolved interval, changing a threshold prospectively and offering an alternative assessment produce different populations. The assistant records the instruction and evidence; qualified owners retain candidate decisions and legal interpretation.

A review deliverable should make uncertainty countable. Report applicants with confirmed old configuration, confirmed new configuration, unresolved version, technical non-completion, accommodation routing, withdrawal and missing disposition. Do not combine unknown with either outcome group. This layout lets qualified reviewers decide what further analysis or candidate action is warranted while preventing an administrative dashboard from presenting an unexplained stage difference as a finding about job relatedness, validity or discrimination.

The change log must also preserve manual exceptions made before and after the configuration change. An override may reflect a documented review, a technical workaround or an unexplained intervention; those states should not be merged. Showing their timing and authority helps the qualified reviewer distinguish tool behavior from later human action without asking the assistant to judge whether either was proper.

Boundary, limitations and conclusion

The assistant may inventory versions, preserve approved criteria, reconcile notices, route accommodation requests, assemble aggregate counts and flag missing approvals. Employment, legal, accessibility, data-science, HR and hiring owners decide job relatedness, validation, adverse impact, accommodation, vendor suitability and candidate outcomes.

No applicant data, employer procedure or vendor model was examined. Counts alone cannot establish causation, fairness, validity or legal compliance; applicable duties depend on the role, employer, tool, population and jurisdiction.

The practical conclusion is narrow: define one screening configuration period linked to job, criteria, tool version, effective time, input fields, threshold, population, accommodation route, human-review rule, approver, notices, outcomes, exceptions and superseding configuration; preserve source, decision and destination evidence; and keep owner-only judgment outside the assistant lane. The reader gets a version-control and denominator design that makes silent configuration drift inspectable while keeping legal, statistical and employment judgment outside the assistant lane.