Best fit
A hiring team can produce a fast AI-assisted shortlist but still have no evidence that the screener is stable across repeat runs, supports negative decisions with evidence, avoids sensitive proxies, or preserves a human-review path.
// Business & Operations
Hiring Screener Audit Harness
Test an AI or rules-based resume screener for outcome changes, score drift, missing evidence, sensitive-term leakage, and human-review gaps.
What problem does Hiring Screener Audit Harness solve?
A hiring team can produce a fast AI-assisted shortlist but still have no evidence that the screener is stable across repeat runs, supports negative decisions with evidence, avoids sensitive proxies, or preserves a human-review path.
Use it to
- Repeat the same synthetic case to detect unstable outcomes or material score drift
- Test identity-varied synthetic pairs without submitting real applicants
- Find negative decisions that have no cited evidence or human-review route
- Create a reviewable audit memo before a pilot or procurement decision
What it returns
- Controlled test plan for repeated and synthetic identity-varied screening runs
- CSV findings for outcome changes, score drift, missing evidence, and human-review gaps
- Severity-ranked audit summary with READY FOR HUMAN REVIEW or DO NOT RELY guidance
- Known unknowns and follow-up tests without a legal-compliance claim
A representative input and result
input case_id,run_id,input_fingerprint,outcome,score,evidence,human_review,sensitive_terms
case-07,run-a,sha256:abc,advance,82,"5 years relevant operations experience",yes,
case-07,run-b,sha256:abc,reject,61,,no,age
result Recommendation: DO NOT RELY
Critical: case-07 changed from advance to reject with the same input fingerprint.
High: score drift was 21 points; the negative result had no evidence and no human-review path.
Medium: decision evidence contained a configured sensitive term.
Next step: isolate the variable, rerun the controlled case, and require human review before any live use.
Access and approval boundaries
- Read access to user-supplied synthetic test plans and screening-run CSV files.
- Write access only to the selected local audit output directory.
- No ATS, hiring platform, network, applicant-record, or employment-decision access is required.
Known limitations
- Results are only as representative as the controlled fixtures and screener outputs supplied.
- Synthetic counterfactual tests can reveal concerning differences but cannot establish why a model behaved differently.
- The audit is operational quality assurance, not legal advice, compliance certification, or proof that discrimination is absent.
// choose with context
Is Hiring Screener Audit Harness the right skill?
It returns
Controlled test plan for repeated and synthetic identity-varied screening runs. CSV findings for outcome changes, score drift, missing evidence, and human-review gaps.
Do not use it as
It does not rank real applicants, recommend who to hire, or make employment decisions.
Compatibility, access, version, and licence
- Reads or accesses
- Declared local scope: screening-audit-input/**, screening-audit-output/**, *.csv, *.json.
- Compatibility
- CSV exports from AI, ATS, rules-based, or human-assisted resume-screening workflows. Python 3 for the included deterministic local audit script. Any screening system that can expose case IDs, outcomes, scores, evidence, input fingerprints, and human-review status.
- Version
- 1.0.0
- Licence
- Review the package licence and seller terms at the linked destination.
Questions
Does this rank applicants or recommend who to hire?
No. It audits the behavior and review controls of a screening system using controlled outputs; it does not score real applicants or make employment decisions.
What data should I test with?
Use synthetic or properly authorized test cases. Include repeated identical cases and carefully controlled variants so a difference can be interpreted.
What failures does it detect?
The included audit detects changed outcomes for identical fingerprints, material score drift, negative results without evidence, missing human-review paths, and configured sensitive terms in decision evidence.
Does a clean result prove the screener is unbiased or compliant?
No. A clean fixture run is limited evidence about those fixtures. Broader validation, governance, legal review, and ongoing monitoring may still be needed.
Does it connect to my ATS?
No. Version 1.0 audits a local CSV export so the test remains inspectable and does not require platform credentials.
Can I change the score-drift threshold or sensitive-term list?
Yes. The helper accepts a drift threshold and an explicit sensitive-term list so the test can match a documented review plan.
Hiring Screener Audit Harness keeps proof and approval boundaries visible.
The listing includes the tested package, realistic samples, declared permissions, and known limitations.
Get Hiring Screener Audit Harness on Agensi