How AI interview scoring works, step by step

What is the rubric an AI interview is scored against?

A scorecard, which Recagn calls a rubric, lists the hard gates, the weighted criteria and the cut-offs a role is scored against. Recagn drafts it from the client's brief, with interview questions linked to each criterion. You edit it and lock it, and every candidate is then scored against that same locked version.

A rubric has four parts:

• Hard gates: rules a candidate passes or fails.
• Scored criteria: each has a weight and a 1 to 5 scale, with a written description of what the evidence looks like at each point.
• Cut-offs: the two scores that separate Qualified, Maybe and Disqualified.
• Questions: each is linked to the criteria or the gate it gathers evidence for.

The example on this page is a rubric for a Senior Java Developer in Pune. It has four criteria and one hard gate, Notice period ≤ 30 days. The client's brief only said "immediate joiners preferred", so a person set the limit.

If you edit a locked rubric and lock it again, Recagn saves a new version. Locking it unchanged does not. Each candidate's evaluation uses the version that was current when they finished the interview, and later edits never change it.

Locking the rubric is one step in how Recagn works, from client brief to your decision.

How are the criteria weighted?

Recagn gives each criterion a weight, and its share of the score is its weight divided by the total of all weights. In the example the weights are 30, 25, 15 and 30, out of 100, so Kafka or RabbitMQ in production counts for 15% of the score. Recagn drafts the weights, and you set them before you lock the rubric.

The example's four criteria and their weights:

• Spring Boot services in production, 3+ years: 30
• Owned a payments or ledger service: 25
• Kafka or RabbitMQ in production: 15
• Code review and mentoring: 30

Weights are relative, so they don't have to add up to 100. Recagn divides by whatever they add up to, which is why the formula ends in ÷ Σweight.

The weights decide the order of the shortlist. That makes them the part worth checking before anyone is invited.

How does Recagn score the answers?

Recagn scores each criterion from 1 to 5 against the anchors written for it, using only the answers to the questions linked to that criterion. Each score comes with the scorer's written reason, and the answers and transcripts sit beside it. You read the evidence and decide whether the score holds.

The scorer is given the job description, the locked rubric and the candidate's answers. The transcript it reads carries no name or contact details. It is told not to use or infer protected attributes such as name, age, gender or photo.

Under Recagn's draft wording, Ananya R. scores 4, 2, 3 and 4 on the four criteria.

Her 2 shows what thin evidence looks like. Criterion 2 asks whether she owned a payments or ledger service. Her answers show she fixed payment reversals while on call, but not that the service was hers. The scorer's reason says ownership isn't shown, and it isn't disproved either. The 2 still counts toward her total.

When an answer can't be read

Sometimes a voice or video answer never transcribes, or the text of an uploaded document can't be extracted. You can then choose to score the interview without that answer. Recagn treats it as no evidence, never as a weak answer.

A criterion that only that answer tested shows "Insufficient evidence" and is left out of the total: its score and its weight drop out of both sums. Wherever the evaluation appears, the Company Report included, it carries a flag such as "Scored without 1 answer (transcription failed)". Recagn never makes this choice for you, and it can't be undone.

What the score doesn't use

Recagn's Authenticity Detector looks for signs that an answer wasn't the candidate's own. It runs as a separate check. Its 0 to 10 reading sits beside the score and never feeds the score or the bucket. You decide whether a concern matters.

How is the weighted score worked out?

Recagn multiplies each criterion's score by its weight, adds the results and divides by the total weight: Σ(score × weight) ÷ Σweight. For Ananya R. that is (4×30 + 2×25 + 3×15 + 4×30) ÷ 100 = 3.35 out of 5, worked out by Recagn's code, not by the AI model. The total and the bucket it gives are a recommendation, which you approve or override.

Each criterion adds score × weight ÷ Σweight to the total, and Σweight is 100 here. Her four contributions are 1.20 + 0.50 + 0.45 + 1.20 = 3.35, the four segments on the ruler above.

The AI model returns only the score for each criterion and the result for each gate. Recagn's code works out the weighted total and the bucket from those by a fixed rule, so the same scores always give the same total and the same bucket. Whether the scores behind it hold is still your call, and you can override the bucket with a note.

Now change criterion 2 to the wording you might write after the client call: "Handled payment reversals in production". Her Q3 answer now shows it. The score is 4 and the total is 3.85. The interview is the same; the rubric is worded closer to what the client wanted.

What is a hard gate in an AI interview?

A hard gate is a pass or fail rule, such as Notice period ≤ 30 days, and failing one makes Recagn recommend Disqualified, whatever the total. Recagn checks the gates before the scores, and counts a gate as failed when the answers don't show it is met. You can override that recommendation.

In the example, Ananya's answer to Q6 is "I can join in two weeks." That meets the 30-day gate.

Two cases are easy to confuse:

• The answers don't settle the gate. Recagn counts it as failed.
• The only answer that tested the gate couldn't be read, and you chose to score without it. The gate is Unresolved. It neither passes nor fails, and it holds the recommendation at Maybe or below.

Either way, Disqualified is a recommendation. It doesn't reject the candidate; that stays your decision.

How does the score become Qualified, Maybe or Disqualified?

Recagn compares the weighted score with the rubric's two cut-offs: in the example, Qualified starts at 3.5 and Maybe at 2.75, so Ananya's 3.35 is Maybe. Below 2.75, or with any failed gate, the recommendation is Disqualified. Each rubric sets its own cut-offs, which Recagn drafts and you edit before you lock.

Recagn applies these rules in order:

1. A failed hard gate gives Disqualified.
2. Otherwise, if every criterion was left out, there is no total, and the bucket is Maybe.
3. Otherwise, a score at or above the Qualified cut-off gives Qualified, at or above the Maybe cut-off gives Maybe, and below it gives Disqualified.
4. An Unresolved gate turns a Qualified into Maybe.

The dashed lines on the ruler are the two cut-offs. With criterion 2 reworded, her 3.85 crosses 3.5 and the bucket becomes Qualified.

The figures 3.5 and 2.75 belong to this example rubric. They are not a default that every rubric uses.

Who makes the final decision?

You make the final decision: you approve Recagn's recommended bucket or override it with a written reason, then set Selected, Maybe or Rejected. Recagn's Disqualified moves no one out of the pipeline, because only your Rejected does. Recagn recommends. Only you can reject a candidate.

Your decision is called a Disposition, and Recagn won't let you set one until you have approved or overridden the bucket. An override needs a note. Only a Disposition of Rejected moves a candidate out; Recagn's Disqualified never does.

Recagn records both: the bucket at the moment you decided, and whether you agreed with it. The gap between the two is kept.

Candidates are told this before they start: see what candidates see and agree to.