Coaching the round

The problem it solves

A round is finite. Left to itself, it re-asks whatever the interviewer finds interesting — which is usually what the last round already established — and the rubric's quiet corners stay quiet forever. Four rounds can produce four overlapping conversations and still leave the criterion that decides the hire untouched.

And nobody tells interviewers how they did. Interviewing is a skill practised constantly and coached almost never, because the feedback would have to come from someone who sat in the room.

Three things address this: a brief before, a co-pilot during, and a critique after.

The brief, before

Before a round starts, the system reads the cumulative coverage map for that candidacy and writes down what this round has to find out:

• Priorities — criteria nothing has evidenced yet, ordered by the weight they carry.
• Contradictions — criteria that are covered but are not settled. Four things put a criterion here: two rounds that scored it far apart (the mean hides a real disagreement, which is why the coverage map keeps a score per round), a score below the retest bar, evidence that is all paraphrase with no transcript moment behind it, and a deal-breaker only one round has read. A deal-breaker decides the hire, so it gets a second, independent read however well the first round scored it.
• Nothing left to cover — when earlier rounds have settled every criterion, the brief says so and tells the interviewer to run the round on their own questions. The round is still scored against the whole rubric.
• Probe lines — a suggested angle per criterion, templated by the kind of requirement.

Three decisions are worth knowing:

The brief is not written by a model. Everything it says is already a fact in the coverage map: which criteria have a score, what they weigh, what was quoted. A model restating that would slow the moment an interview starts, could fail when it must not, and could invent a criterion that is not in the rubric.

The first round gets a brief too. There is no coverage map until an interview has been scored, and a brief that said "no data" would be useless in exactly the round that sets the tone for the rest. With no coverage map, every criterion in the rubric is uncovered — which is true, and is the most useful thing the brief can say.

Résumé gaps are first-round targets only. What the application failed to establish is the best question list that exists before anyone has spoken to the candidate, and the worst one afterwards. By round two the interviews themselves know what is still uncovered, and re-asking what the paper missed spends the round on ground round one already took. So the résumé's gaps are read only while the coverage map is still empty. A brief written after a fallback résumé score carries none, because the deterministic scorer produces no gaps.

Starting the same interview twice refreshes its brief rather than writing a second one.

The co-pilot, during

In the room, the co-pilot suggests the next question. It does call a model — but it can only ever aim at a criterion the brief already named, and the criterion it names is checked against the brief before the suggestion leaves the server. A suggestion targeting a criterion that is not in the rubric is worse than no suggestion.

Given the live transcript, a suggestion is a probe drawn from one of four places, and it says which:

• A contradiction: something the candidate just said conflicts with what they said earlier, in this interview, in their résumé or in an earlier round.
• A résumé gap: the résumé or the screening notes leave something open.
• An uncovered criterion: a rubric criterion no round has evidenced.
• An unanswered question: a question was asked and not properly answered.

The first and the last are claims about what was said, and a model will invent them. So they carry quotes, each quote is checked against the text it claims to come from, and a probe whose quotes are not there is discarded and replaced by one built from the brief. A contradiction needs two checked quotes, one of them the candidate's. An unanswered question must quote the question from this interview.

The critique, after

After a human-led round, the system writes one review of the interviewer, covering four questions a hiring manager actually asks:

• Rubric coverage — how much of the job's rubric this round managed to evidence.
• Question quality — were the questions specific, probing, and followed up.
• Bias and leading — were any questions leading, biased, or out of bounds.
• Talk ratio — how much of the conversation the interviewer took up.

These critiques are visible to admins and hiring managers. That was chosen knowingly, and the mitigation chosen with it is that every claim carries a quoted example from the interviewer's own turns. A critique point the model cannot quote verbatim is dropped on the way in, and a judged dimension with no surviving quoted point gets no score at all. Nobody's manager ever reads an unevidenced sentence about their competence.

Rubric coverage and talk ratio are exempt from the quoting rule, and only those two: they are arithmetic over stored rows rather than assertions, and their rationale says exactly what was counted.

One person is reviewed per interview: the participant marked as the round's lead. The lead sets the questions and owns the round, whereas a shadow or an observer may not speak at all — attributing a panel's talk ratio or its leading question to a panelist who said two sentences would be precisely the unevidenced claim about a named colleague this design exists to prevent. An interview with no lead is not reviewed.

What the user does

• Read the brief before joining. It is short by design.
• Use the co-pilot's probes, or ignore them — nothing depends on accepting one.
• Read your own reviews. Managers read their team's.
• Treat the contradictions list as the thing worth spending the round on, since it names disagreements the average score is hiding.

Limits and edge cases

• The interviewer review covers human-led rounds only. The AI screen's equivalent is the never-ask transcript scan, which is a compliance control rather than a coaching one. See Compliance.
• The co-pilot needs you signed in. It reads the job's rubric, the candidate's coverage, their résumé and earlier rounds' transcripts.
• Talk ratio can be unavailable. If the transcript's speaker labels are not recognised, the ratio is reported as unavailable rather than computed wrongly.
• The brief is only as good as the coverage map. A round that produced no transcript contributes nothing, so the next brief will still list its criteria as uncovered — which is accurate.
• A critique is per round, not per person over time. Each review is a judgement about one conversation.
• Nothing enforces the brief. An interviewer is free to ask something else, and the round is scored on what was actually said.