A Tab Switch Count Is a Flag, Not a Score Change

A hiring manager asked me, in one line, whether three tab switches drop a Go to a No Go. They do not. A tab switch count can stop the live call, and it can light a flag after the call. It is not allowed to move the score. We split those on purpose, and the split is the part teams blur the first time a strong resume dies mid-sentence.

If you only remember one thing, remember the two machines. The first machine sits in the browser during the call. It shows Tab Switch Detected three times and ends the interview on the next hide. That machine cares about whether the conversation got to exist. The second machine runs later, over the finished record. It may mention page-hides as one input among others, under a band called low, elevated, or high. That machine cares about whether a person should listen before trusting the words. Neither machine edits the points.

The live warning is described in what happens when Tab Switch Detected appears. The flag is the rest of this page.

Key takeaways

• Ending the call and scoring the call are separate steps.
• The AI-assist panel says signals, not proof.
• Tab switches are worth 15 to 35 points, capped, so they cannot reach the high band alone.
• Across 1,447 production interviews, 3.4% showed any tab switch.
• A short cutoff is an incomplete interview, not a secret fail.
• If you want a human penalty, write it as a human decision. Do not pretend the score already did it.

Two machines, and why the score stays still

The live machine is blunt because it has to be fast. The page hides, the count ticks, three warnings, the fourth hide disconnects the room. There is no model in that path. There is no “this hide looked innocent.” Fast and blunt is acceptable only if the consequence stays small enough to repair: the person may need another slot, and whatever they said is still judged on what they said.

The later machine is slower and more cautious, and it still does not touch points. After the call, HireQwik computes AI-assist signals from things the interview already recorded: timing, speech patterns, and the proctoring counts. The panel on the candidate record is titled AI-assist signals. The line under the band says these are clues rather than proof, and that you should watch the recording before you act. I keep quoting that line because teams crop it when they screenshot the band into Slack.

Nothing downstream is supposed to read that band. Scores, verdicts, and automatic reject rules are computed without it. If someone wires the band into the score later, the tests we keep on that boundary are meant to fail. I will take one extra human conversation over a hidden penalty. An hour of follow-up is annoying. A mark the candidate cannot see, and cannot argue with, is a different kind of damage.

So when a recruiter says “the score was dragged down by tab switches,” they are describing a step we did not build. Correct it in the room. The score came from the answers and, on communication, from how those answers were delivered. The count is sitting beside that, as a flag, or it ended the call before a full score existed. Those are the only two stories.

How the tab switch count is weighted, and why the cap exists

Inside the flag, not every input is equal. Page-hides are one of seven. The weight for this one is simple: 15 points for the first switch, plus 10 for each further switch, and the total for this input stops at 35. One switch is 15. Two is 25. Three or more is 35. It does not become 45 because someone left the page six times.

The bands are fixed. Elevated starts at 35. High starts at 65. A capped input of 35 can open the elevated band by itself and can never open the high band by itself. High means several different signs agreed. A single odd behaviour, even a repeated one, is too easy to explain with ordinary life. You want the browser, the clock, and the voice pointing the same way before you treat the row as urgent.

The other inputs, so you know what agreement looks like, are a long wait before answers, a pattern we call slow-then-fluent, speech that sounds read, essay-style spoken structure, a long stretch of talk with no repairs of their own words, and video disabled more than once. I am not re-teaching each of them here. The one that catches live chatbot use, the long wait followed by oddly smooth speech, is the centre of how we look for ChatGPT during the interview. A candidate who was reading a prepared script, rather than waiting for a tool, has a different sound, covered in what a read-aloud answer looks like. Page-hides are the browser’s contribution to that same panel, not a private court.

Here is a worked sum, so the cap is not abstract. Three switches are 35. Add slow-then-fluent at 25, and a long call with zero self-repair at 10, and the total is 70, inside the high band. Three switches and nothing else stay at 35, the floor of elevated. Same count, very different ask of the reviewer. I would open the recording the same day for the 70. For the 35, I would listen only if I was already unsure, and I would not build a case on it.

Why we refused to let the count move points

We had the numbers to build a naive penalty, and they talked us out of it.

Before the flag shipped, we looked at 1,447 completed production interviews from a 60-day window. Any tab switch at all showed up in 3.4% of them. That is uncommon, which is why it is allowed to count for something. It is not rare in the way a confession is rare. Three or four in a hundred will include people finding headphones, people whose laptop slept, and people who were, yes, somewhere else. A rule that subtracts a point from every one of those records would be a tax on 3.4% of candidates, collected without a hearing.

Timing taught us the same lesson in draft form. A cutoff of eight seconds of silence before an answer felt strict. On that 1,447-interview set it would have painted roughly thirty percent of ordinary candidates as suspicious. The middle of the pack waited 6.9 seconds, while one in ten waited 15.1 seconds and up. Nerves, care, and a second language all live in that spread. A penalty sitting in the middle of normal behaviour invents fraud. It does not find it. The tab-switch cap is the same instinct: stay at the edge, and even there, keep the edge off the mark.

We then replayed the whole flag on 2,507 older interviews. Almost all of them, 97.4 percent, landed low. Forty-five were elevated. Two were high. Eighteen could not be judged at all. I trust a light that stays dark nearly every day. A light that trains the team to shrug by Friday is useless. Leaving it off the score is what makes people willing to look on the day it comes on.

Harvard Business Review has asked the uncomfortable version of this question in public: whether you are interviewing the candidate or their AI. The honest operational answer is that you often cannot know from text. You can collect a few hard signals and hand them to a person. Handing them to a formula that subtracts marks is how you get a number nobody can explain in a debrief.

What you see on the record, and how to use it

On Interview Details, when the band is not the empty “insufficient” result, you see the band name, a 0 to 100 figure that belongs to the flag and not to the interview, a count of inputs examined, a count of inputs that fired, and a line telling you this is a clue and that you should watch the recording before you act. Open the breakdown and each fired input shows its points. Page-hides appear there, with their weight, when this input fired.

Read the band before the list. The band tells you whether the case is rare or mild. The list tells you whether you only have page-hides or page-hides plus the way they spoke. Then play the recording at the moment the list implies. For page-hides, the silence after the question matters more than a hide while the question was still being asked.

Insufficient is not a clean pass. The call had nothing usable: no timings, no speech measures, no proctoring payload, no transcript. Treat a blank band like a reference that never arrived. On a borderline interview, listen, rather than taking comfort from a blank band.

If the live machine already ended the call, you may be looking at a short record and a flag together. Do not stack them into a story the product did not tell. The end means the conversation stopped. The flag means the pattern, on whatever audio exists, looked worth a listen. A two-minute stop often will not have enough speech for the interesting audio signals at all. You are back to the incomplete rule, which is a process decision: rebook or close. The cutoff decision is that decision. The flag does not make it for you.

The human penalty, if you still want one

Some teams want a penalty anyway. I understand the impulse. A candidate who left the page four times and still produced a tidy Go feels like the rubric got played. Maybe it did.

If you believe that, make it a human decision in the review queue, with the recording on. Write “live follow-up, page-hides plus read-sounding answers” on the record. Invite them to a conversation with a person. See if the tidy answers survive a follow-up they cannot pre-write. That is a penalty with a face on it. It can be reversed if you were wrong.

Do not encode “minus one band per warning” in a spreadsheet beside the official score. Within a month you will have two scores, they will disagree, and the unofficial one will be the one a manager quotes because it is harsher. You will also apply it unevenly, because the person who remembers the rule on a Thursday will forget it on a Monday. The product’s refusal to do this automatically is not a missing feature. It is the decision.

There is a related temptation on campus drives: auto-reject every terminated row so the cell sees a clean list. That list is clean because you deleted the hard cases. The hard cases are exactly the ones a placement office will ask about. Keep them visible, label the next action, and move. The drive-day version of that labelling is in campus tab switch warnings.

What a changed score would have to mean, and why we will not imply it

If we ever let this count move points, we would owe every candidate a visible line: this many hides, this many points, here is the appeal. We do not have that line, because we do not have that penalty. Anyone on your team who talks as if we do is making policy in the corridor.

The things that do move the interview score are the answers against the role’s dimensions, and the delivery blend on communication when the audio is good enough to blend. Poor audio has its own path, and it is not this one: a noisy or silent recording can change how communication is judged, which we spelled out in how bad audio is handled. Page-hides are not audio quality. Do not file them in that bucket because both feel “technical.”

A candidate who finished the call with one early hide should be read as a person who finished the call. A candidate who never finished should be read as unfinished. A candidate whose flag is high should be listened to before you trust the prose. Three sentences. None of them is “subtract two points.”

If your review meetings are still treating the band as a mark, book a demo and bring the person who runs those meetings. We will open a record, show the interview score next to the flag, and agree on a sentence your team can repeat without inventing a penalty.