Among 139,336 candidates that Metaview scored in both an early round and a final round, the two assessments pointed the same way 54.4% of the time. Put another way, on nearly half of these candidates, the early read and the later read reached different conclusions. That does not mean the early interviewer was wrong whenever the two disagreed. A later round often tests different skills, gathers more evidence, or brings in someone with a different view of the role. What the number does show is that candidate evaluations change a great deal as a process runs its course.

That matters because most candidates do not get several chances to be assessed. Only 20.8% of the candidate and job pairings in Metaview's corpus included a second recorded round. For the other four in five, one early read carried most of the weight. So the finding worth sitting with isn't that first rounds are unreliable in some absolute sense. It's that early assessments change often and still decide most outcomes, frequently on thin and undocumented evidence.

This data can't tell you how many qualified candidates were wrongly rejected. The people cut in an early round rarely come back with a job-performance outcome to check against, and, as you'll see, they barely appear in this dataset at all. What the data can do is map where that risk is highest: early decisions made with inconsistent criteria, on incomplete evidence, with no written record. Those conditions are measurable, and they're fixable.

Early and final assessments often disagree

Here is the finding in full. Metaview looked at 139,336 candidates who received a scorecard in an early round and another in a final round, then checked whether the two recommendations pointed the same way. They matched 54.4% of the time. Excluding the neutral, no-strong-opinion scorecards moves it only slightly, to 55.2%. Either way, the same candidate is assessed differently at the start and the end of a process close to half the time.

These figures come from Metaview's own corpus of 5.5 million captured conversations across 12,491 organizations, pulled on June 17, 2026, of which 5.2 million are candidate interviews. Everything here is aggregate only, with a floor of at least 50 interviews behind every number.

What that number does and doesn't say

It's worth being precise, because this number is easy to overstate. Agreement isn't the same as accuracy. When an early and a final assessment disagree, the data doesn't say which one was right. The later round may simply have tested different competencies, seen more of the candidate, or applied a different bar. Disagreement tells you the reads were inconsistent, not that either was a mistake.

The sample matters too. To appear here a candidate needed both an early and a final scorecard, which means they reached a final round. So the 139,336 skew toward candidates whose early read was positive, mixed, or overridden. It only includes candidates who advanced far enough to be scored twice. That is a real limitation, and it points straight at the harder question underneath all of this.

What you can say is that early assessments are inconsistent, and wider research has found the same for years. A classic meta-analysis by Conway and colleagues put the inter-rater reliability ceiling for unstructured interviews, still the default first-round format, at 0.34, meaning two interviewers watching the same candidate often reach different conclusions. Structured interviews roughly double that. None of this measures Metaview's 54.4% directly. It describes the same underlying pattern: unstructured early screening is not a consistent instrument.

A first-round no isn't always the final word

Sorting outcomes by what the scorecard recommended shows the early call isn't always binding. Candidates marked NO still advanced 6.2% of the time, against 47.4% for those marked STRONG_YES. So a NO recommendation lowered a candidate's odds a lot, but didn't always end the process.

Be careful reading anything into that 6.2%. Advancing isn't the same as being hired or performing well; some of those candidates were surely turned down in the next round. And the data doesn't say why any single override happened. It could reflect genuine disagreement between interviewers, new evidence in a later stage, or a recommendation that was never treated as decisive. What the figure does show is modest but real: the line between a first-round yes and no is softer in practice than a clean sequence of gates would suggest.

Half the time, the scorecard is missing

When Metaview looked at candidates who advanced, 41.9% had no submitted scorecard on file at the moment the decision was made. That's a genuine operational gap, though it's worth stating carefully what it means: the basis for the decision isn't visible in this dataset. It doesn't prove no evidence existed. A scorecard might have been filed later, feedback might live somewhere else, or a recruiter might have moved a candidate forward before all the written feedback landed.

Relatedly, in multi-interviewer loops, 56.3% of the candidates who advanced did so with at least one dissenting no on record. On its own that isn't alarming. Panels are partly designed to surface independent, sometimes conflicting views, and a single no among several yeses can be a healthy process rather than a broken one. To know which, you'd need to see how many interviewers were involved, whether the dissent touched a core competency, and whether it was resolved in the debrief. It's a reminder that one strong opinion, for or against, is rarely the whole story.

The pattern these gaps point to is one a lot of hiring research warns about. Google's structured interviewing guidance describes how a snap first impression pulls interviewers toward confirming their initial read. Harvard Business School's Hidden Workers study found that 88% of employers believe qualified candidates get filtered out of their process for not matching the exact criteria in a job description. Neither study measures Metaview's numbers. What they establish is the mechanism: inconsistent, under-documented early screening is where qualified people quietly fall out.

Where the risk is highest

So here's the honest version of the stakes. This data cannot count how many qualified candidates were wrongly rejected. It can identify the conditions in which that risk concentrates: early decisions made with inconsistent criteria, on incomplete evidence, with no written record. A rejection made under those conditions is the kind most likely to be a mistake, and the least likely to be caught, because there's nothing to review. That is a narrower claim than a headline false-negative rate, and it's the one the numbers actually support.

The most reliable lever is also the least glamorous. Structured interviews predict job performance better than unstructured ones, 0.51 against 0.38 in the coefficients Schmidt and Hunter established and the U.S. government still cites. They work because they run every candidate through the same questions and the same bar, which is precisely what an improvised early round does not do. Consistency is the whole point.

Structure only helps if the reasoning survives the conversation, and most of it doesn't. An interview happens, a few notes get typed, and the detail is gone by the debrief. That's the gap Metaview was built to close. Its Notetaker captures every spoken word of an interview, so the basis for a decision is a record you can go back to rather than a memory you reconstruct.

How to make the first round more consistent

You can't fully measure false negatives, but you can shrink the conditions that create them. Three moves, in order of impact.

First, make sure everyone gets a real review. A common way strong candidates fall out early isn't a bad decision, it's an unmade one: a human starts at the top of the stack, works down until the calendar runs out, and the rest go unread. Metaview's Application Review reads every inbound application against the ideal candidate profile you define, sorts them by fit, and shows its reasoning on each, so a candidate deep in the pile gets the same read as the one who applied first. There's more on how it works in our guide to inbound screening, but the guardrail is the important part: it never auto-rejects and never sends a rejection on its own. It reads, ranks, and explains, and a human decides who moves forward.

Second, make the first round leave a record. A rubric turns a vague impression into criteria you can score, and a captured interview means the score is backed by what the candidate said. A tight structured scorecard and a shared bar, the kind a bar raiser program sets, turn an early no into a reason you can review and compare rather than one you take on trust. Here's how one talent team describes the shift.

Third, measure your own consistency instead of trusting a benchmark. Metaview's Reports lets you ask your pipeline in plain language how often your early reads match your final decisions, and which interviewers, stages, and roles are the least consistent, so you can tighten the rounds that need it. It's the same idea behind agentic recruiting generally: capture the work, then surface the pattern inside it.

None of this asks a recruiter to work harder on the queue. It asks the process to stop losing the signal it already generates. If you want a wider view of the category first, our guide to the best interview intelligence tools is a good next read.