Screening

Every application is read by a model against the job's requirements. The read is advisory: it decides nothing. Concepts describes what it produces — a score, a summary, a line per requirement and a short list of findings, each claim carrying the quote it came from.

What the model is given#

Every read is sent:

• The job's title, requirements and description.
• The labels of the job's form questions, and the candidate's answers to them.
• The cover letter.
• The résumé — the PDF itself when it is at most 2 MiB and 30 pages, otherwise the text extracted from it.
• The candidate's name and location.

The candidate's email address and phone number are not part of the prompt, though a résumé usually contains both, and the résumé does go.

Only a PDF résumé reaches the model. The apply form accepts DOCX too, but nothing is extracted from one, so a DOCX candidate is read on their answers and cover letter alone and their read can quote nothing from the résumé. If evidence looks thin for someone, check what they uploaded.

Every attempt is recorded in an append-only audit entry — the model id, the rubric version, what was shown, what came back — which stores the résumé's storage key rather than its bytes. See Data and privacy.

Re-reading when new evidence arrives#

Settings › General › Screening › Re-reading is the one dial a workspace has. It is on by default.

Re-reading	What happens
On	A submitted scorecard, a reply from the candidate or an interview outcome prompts a fresh read, which is also sent the read it revises and the evidence that arrived since. Each one costs a model call.
Off	A read is made once, when the application arrives. Re-read on the packet is the only way to ask for another.

A read is retried at most three times. When all three fail the packet says so, and Re-read starts the count again.

What candidates are told#

One sentence appears on every public apply form, above the submit button:

We use AI to help review applications against what this role asks for. A person reads every application and makes every decision.

The same fact is repeated in the application-received email, together with how to see or delete what was shared, and under every read on the packet: "rubric v3 · <model id> · Reviewed with AI assistance. A person makes every decision".

You cannot edit any of it out. Settings shows the exact wording, and the model id it is read with, read-only.

When no model is configured#

None of that notice appears: it is shown only when a model is configured, and Settings says so too.

Everything except the read still works: applications arrive, stages move, mail sends, interviews book. The score column shows ·· instead of a number.