The Approval Click: What Live-Production Review Really Catches

A highlight clip goes to air after a person says yes. What would it take to say how often that yes had to be earned?

Andreas EhstandA common design for automated highlight production puts a person behind the machine. The system proposes a clip, a caption, a score graphic; an operator approves it before anything goes out. The arrangement is easy to describe and reassuring to hear. This Op-Ed looks at one narrow question about it: what the approval, as a number, could and could not tell us.

This is an observation from outside, written as a reading aid. It makes no statement about any vendor, product, or broadcaster, and it does not replace a broadcaster's editorial standards, its contractual requirements, or the validation of a particular production system. All figures are constructed for the illustration.

What the Click Records

An approval is a single bit. It records a decision; by itself, it does not establish what the operator checked. An operator comparing the caption with the roster and an operator glancing at the thumbnail can leave the same trace in the log. A high approval rate can reflect accuracy, permissive review, or both. The rate alone cannot say much about the review. A count against errors whose location is known in advance can say more.

A Constructed Rehearsal

Imagine a rehearsal on the replay of an archived match. The system's proposals for 40 clips are prepared. Before anyone sees the queue, a colleague who is not the operator plants errors in five clips and writes down which ones: a wrong name in a caption, a score graphic that is one goal off, a clip from the wrong half, a swapped team label, a shirt number that does not match the name. The operator then reviews the queue at a pace meant to approximate a live broadcast.

In this constructed case, four of the five altered clips are held back, and three of the 35 untouched clips are held back as well. The run yields two figures, and each answers its own question.

The first, four of five, says how much of the known material the operator caught. The second, three of 35, says how many of the unaltered clips were held back anyway. A queue in which everything is held back would catch every planted error and could not be used on-air. A queue in which almost nothing is held back would raise few flags, and holding back few clips does not, by itself, establish how many errors were detected. Reporting either figure alone gives an incomplete account of this run.

With five planted errors, the first figure is soft. A different result in a second rehearsal could reflect the selected errors, the operating conditions, or chance, and nothing in this single run separates these from a real difference in attention.

What Planted Errors Cannot Say

The altered clips were made by someone who knows what a mistake looks like and chose five that a person can find. The system's own mistakes may look different. A plausible but wrong name, a clip that is almost the right moment, a graphic that lags by a few seconds: errors of that kind may be harder to see than a swapped team label. A catch rate measured on planted errors, therefore, describes the operator against planted errors. Whether it carries over to what the system gets wrong by itself is a separate question, one the rehearsal does not answer.

A Flag Not Planted

Suppose that, of the three untouched clips held back, one really does have a problem, a caption that is subtly off, which nobody in the rehearsal knew about. Filing it under false alarms would penalize the operator for the very thing review is for.

One way to treat it is as a third category: held back, not on the list, waiting for a human decision. If that judgment confirmed one genuine problem and found the other 34 unaltered clips sound, the figure for needless flags would be two of 34 rather than three of 35. The count then becomes a count after judgment, and the judgment becomes part of the record. It also means the list of known errors can grow from one rehearsal to the next, which makes later counts harder to compare with earlier ones.

Speed Belongs to the Design

Some live workflows add a condition that a desk review does not have: decisions are made within seconds, with the next clip already waiting. A count taken at leisure describes the operator at leisure. A rehearsal at live pace is closer, but whether rehearsal pressure resembles the pressure of a broadcast is, again, open. The number is a statement about that rehearsal: that operator, that queue, those planted errors, that pace.

What a Number Would and Would Not Say

A catch rate from a rehearsal reflects this operator working with this interface and queue under these conditions; it does not isolate their separate contributions and is not a statement about other crews or about any particular broadcast. Whether a given production already has counts like these is not something this article can address; it notes only that the approval log by itself does not contain them.

The click alone does not establish review quality; assessing error detection takes separate evidence, and, for some setups, that evidence may already exist. The observation here is only that a person approves it and the approval catches what it should are two different sentences.

Andreas Ehstand is an independent researcher in Starnberg, Germany, and founder of AUGMANITAI. He works on how reliably people who review AI-supported output catch errors whose location is known in advance and how often they flag sound items by mistake.

LINK:	https://www.sportsvideo.org/2026/10/30/the-approval-click-what-live-pr...
	See more stories from svg

More from Sports Video Group