Concepts

Six words worth defining.

Stages#

An application sits in exactly one stage: Applied, Screening, Interview, Offer, Hired, Rejected or Withdrawn. That is the whole pipeline; there are no custom stages. Rejected has no column on the board and is not on the forward path: it has its own control, behind a confirmation, because it is the one move that can email the candidate.

Withdrawn is the candidate's own act, from their page. It is terminal, it cancels any interview that has not happened, and it is not a rejection: it is not counted as one anywhere. Nobody on the team can move an application into it: the stage is a record of something the candidate said, so only they can say it.

The packet#

The packet is the single screen for one application, at /jobs/<job id>/applications/<application id>. What the candidate gave you is on the left — details, answers, résumé, timeline — and what the model made of it is on the right, so it is always obvious which one you are reading. Interviews, messages, scorecards, discussion and activity are tabs underneath.

The read#

The read is what a screening model produced for one application. It is shown the job's requirements and everything the candidate submitted, and answers with a score out of 100, a one-sentence summary, a line per requirement and a short list of findings.

Coverage is that line per requirement, in the job's order, each with a verdict — ✓ met, ~ partly met, ? unclear, ✗ not met — tallied as "N of M evidenced", where evidenced means met or partly met.

Evidence is the receipt under every claim: a verbatim quote, and the answer, résumé page or cover letter it came from. The model writes the quotes, and Krinoa then goes looking for each one in the actual source: a quote it cannot find is dropped, and a claim left with no evidence is demoted — a coverage line to unclear, a finding to a question. Quotes from a résumé with no text layer, a scan, are marked unchecked rather than passed off as verified.

Findings are what no requirement has a line for — a strength, a concern, a question or a flag — and some suggest a next move: Ask drafts a message for you to send, Probe puts a question into an interview round's prep notes.

The read decides nothing: its output has no stage, reject or recommendation field, and no automation moves a candidate. Interviewers do not see it. Screening covers giving your workspace a model to read with.

The thread#

Every application has exactly one thread: the ordered list of messages between your team and that candidate, on the packet's Messages tab. Email and WhatsApp both ride it, so the whole conversation is in one place whatever channel it arrived on. A thread has an owner — the person answering it — which is a different word from an application's assignees, the people meeting the candidate.

Interview rounds#

A job's interview plan is an ordered list of rounds: a name, a duration, a format, default interviewers and prep notes. The round is the template; an interview is one round actually scheduled for one candidate. A round's prep notes are copied onto each interview and go out in the interviewer's calendar invite, which is how a probe raised by the read reaches whoever runs that round.

The ledger#

The ledger is the workspace-wide audit trail: one entry per mutation, with a kind, the actor, the subject and a timestamp, and a null actor where the system or the agent did it (a public apply, or a read). Nothing in the product updates or deletes an entry. You see it as the packet's Activity tab and the feed on Today.