Fairness & Compliance

AI hiring screens are regulated (e.g. NYC Local Law 144, the EU AI Act's high-risk classification, Title VII disparate-impact standards in the US). Kashif is built so you can use AI screening and still defend every decision.

Non-removable fairness rules

Every AI fit score runs under a fixed fairness and safety rule set that is prepended to the scoring prompt outside any editable prompt override. It cannot be turned off or edited away. It instructs the model to:

• evaluate only job-relevant evidence (skills, experience, qualifications, demonstrated ability);
• never consider or infer race, ethnicity, national origin, gender, gender identity, sexual orientation, age, religion, disability, pregnancy, marital/family status, or any protected characteristic — and to ignore names and photos as signals;
• flag proxies (e.g. "recent graduate", "cultural fit" used as a stand-in) rather than weight them;
• never fabricate facts, and separate genuine gaps from merely missing data;
• base every evidence claim on a specific candidate field.

Even if a workspace or operator customizes the scoring prompt, this preamble always leads it.

Deterministic knockout rules

Hard, non-negotiable gates (work authorization, minimum experience, location) belong to humans, not a probabilistic model. Knockout rules on apply questions decline disqualified applicants before any AI call. This removes the "everything rides on an AI score" objection and keeps the gates auditable. Configure them per apply question — see Jobs & careers pages.

Blind screening (optional)

Turn on blind screening under Settings → AI Fairness to hide the candidate's name from the model input only. Your team still sees the full record; only what the AI sees is redacted. Evidence still works because it cites fields, not names. The redaction is reflected in the run's input hash for auditability.

The fairness rules above always apply. Blind screening is an additional, optional layer.

Screening health

Each job's Insights modal has a Screening health section showing how decisions were made: auto-shortlisted, auto-rejected, knocked-out, and % human-reviewed. This is decision-pattern visibility — it reports how the funnel was operated, and never infers protected characteristics.

Compliance pack

For any role, an admin or HR manager can export a compliance pack from the job Insights modal (JSON, or CSV for the decision table). It bundles everything needed to explain and defend the role's screening:

• the job and its calibration,
• the prompt version(s) and fairness posture used,
• every scored candidate's score, confidence, and cited evidence,
• the full stage-transition history,
• consent timestamps,
• the automation settings in force,
• the scoped audit trail.

The export is access-controlled (admin/HR only), audited, and served no-store.

Consent & candidate rights

• Applicants explicitly consent at apply time; the timestamp is recorded.
• Candidates can withdraw at any time from their status page.
• Workspace data can be exported and is subject to retention purging — see Team, roles & security.