Explainability statement

SiftFirst helps a person screen a pile of job applications: it ranks candidates against a rubric the employer sets and shows the evidence behind every score. It is a decision-support tool for a human reviewer.

It does not make hiring decisions and it never rejects a candidate automatically. It does not source candidates, contact them, or take any action on its own. Every decision is made by a person.

The job description, the text of each application, and a weighted rubric of criteria that the employer sets (SiftFirst can suggest criteria from the job post, but the employer keeps, edits, or replaces them).

Each candidate is scored from 0 to 10 on every criterion, and each score comes with a short quote from the application that supports it. The overall score (0 to 100) is a plain weighted average of the per-criterion scores using the employer's weights. Scores sort candidates into a starting shortlist; the employer can re-weight criteria and the ranking updates instantly. The scores are the tool's; the decision is the human's.

SiftFirst scores an application only against the job-related criteria the employer sets. It does not ask for, infer, or use protected characteristics such as race, sex, age, disability, or national origin.

SiftFirst quietly flags applications that read as AI-generated, as a prompt for the reviewer to look closer. It is an internal signal only: it never rejects a candidate, is not shown as a public badge, and does not lower the score. Clean or generic writing is not treated as proof of anything.

The criteria are set by a person, a person reads the quoted evidence behind each score, and a person makes every hiring decision. Saved screenings keep a record of the criteria, the scores, and the evidence, so decisions can be reviewed.

Screenings run signed out are not stored. Signed in, screenings are saved to a private history the account owner can revisit and can delete, along with the whole account, at any time. Data is not sold.

Scoring uses a large language model accessed through OpenRouter (an Anthropic Claude model by default). Model outputs can vary between runs, which is why every score carries its evidence for a human to verify.

A language model can misread or miss things, so the quoted evidence should always be checked; the scores support human judgement rather than replace it. SiftFirst is not a background check and this statement is general information, not legal advice.