Screening Agent: fast yes, accountable no

Prototype for the Darwinbox take-home (D3). The agent answers every frontline applicant within minutes and books strong candidates straight into interviews. A “no” is only ever a published knockout rule the candidate can contest, or a recruiter’s decision. Every step writes a decision receipt. All data is synthetic. Scoring runs in your browser with deterministic rules; in production an LLM extracts the evidence (D2 R4).

Try each route: answer “No” on a must-have, “Not sure” on lifting, or strong answers. Then switch to Germany and repeat.
Warehouse Associate · Night shiftHiring assistant · verified business

Behind the conversation

What the agent did, which rule fired, and the receipt it wrote. This is what a recruiter or an auditor can see later.

Review queue (sorted by likely-yes first · SLA 3 working days)

Country policy (enforced server-side · every change logged)

Approved criteria · Warehouse Associate (v3)

Propose a new criterion

Try “Candidate under 35”, “Graduate from a top college”, “No employment gaps”, “Lives in pin code 421302”, then “Has a forklift licence”.

Decision receipts (append-only · includes those you created in this session)

Select a receipt to inspect it.

Adverse-impact ratio by stage (four-fifths rule · synthetic)

Selection rate of each group ÷ the highest group’s rate. Below 0.80 the rule auto-pauses and the TA admin is alerted (D2 R11). US groups come from voluntary EEO self-ID, which is stored apart from scoring.

Backtest: where last year’s hires would have landed