What Can an Observer Know?

In January 2019, Keith Lyons tried a new R package for looking at missing data.

The result appeared straightforward.

His Asian Cup dataset looked incomplete in the fields for red cards and second yellow cards.

But there was a catch.

The missing values were not necessarily missing observations.

Keith had used NA when a red card or second yellow card had not been awarded.

The visualisation was showing absence in the data structure. To understand what that absence meant, the observer still needed to know the coding rule.

That small example makes a surprisingly large point.

Something can look missing without being unknown.

And that means the central problem of observation is not always whether we can see more.

Sometimes it is knowing what, exactly, we are seeing.

Seeing is already mediated

Clyde Street had been circling this problem for years.

In January 2010, Keith was thinking about television, replay and officiating after controversies in football and cricket.

His concern was larger than whether video should be used to correct a decision.

Television had changed the conditions of seeing.

The viewer at home could pause, replay and inspect an event through mediated images. Spectators inside a stadium increasingly expected large-screen replay. The real-time event and the broadcast event were no longer the same visual experience.

Keith described the televised view as mediated and constructed.

That matters because replay can feel like a removal of uncertainty.

Look again.

Slow it down.

Change the angle.

Surely the truth will become visible.

But the same post carries a more difficult warning. Keith follows work on offside judgement suggesting that both human observers and technological media face temporal constraints when trying to locate several rapidly moving players at the instant a ball is played.

The problem is not simply that one observer is careless and another device is precise.

Different forms of observation encounter different limits.

Keith’s response in that post is deliberately provocative. He wonders whether sport should value the person as arbitrator rather than treat media as judge, and he describes sport as essentially fallible.

We do not have to adopt that conclusion to see the question it opens.

What becomes visible when an event is replayed is not identical to what was available to an official in real time.

The replay may add information.

It may also change the observational task.

The observer is part of the instrument

Clyde Street also contains a much less glamorous version of observation.

On New Year’s Day 2013, Keith published historical data from England–Wales rugby matches he had notated in real time between 1987 and 1992.

The post is unusually useful because it does not present the numbers as though they had appeared by themselves.

Keith tells us how he made them.

He had defined a small set of game events. He used one notation sheet per half. Two stopwatches tracked total time and ball-in-play time. He had operational definitions. He deliberately avoided trying to capture granular detail in real time. He used breaks and half-time to refocus his attention. He was conscious of observer drift and of the possibility of timing error.

He also says something that keeps the record honest.

Keith considered himself a trained observer and was confident that his observations were valid and reliable.

But he had not undertaken intra-observer or inter-observer reliability studies on the data presented in the post.

Both statements belong in the account.

His training and procedures matter.

So does the missing test.

Observation here is not a camera pointed at reality.

It is a designed human procedure.

The observer chooses definitions, granularity, timing rules and attentional limits. They manage cognitive load. They can drift. Their instruments can introduce error. Their confidence can be informed without being independently demonstrated by the particular reliability tests they did not perform.

That does not make the historical data worthless.

It tells us what kind of knowledge those data are.

The procedure gives the observations shape.

The procedural limits give the interpretation a boundary.

Memory is another observational state

A little over a month after writing Observation +, Keith encountered a different kind of observational problem through a Radio National programme on history and memory.

The programme took him back to life-history work and to Ian Franks’s writing on eyewitness testimony.

The Clyde Street post itself is brief. It does not give us a reliability theorem for memory, and it would be wrong to manufacture one from the title.

What it does do is place eyewitnesses, memory and oral history inside the same expanding field of attention.

An event can be observed once and then encountered again through memory.

That later account is not the same object as a live notation, a broadcast replay or a database record.

Each carries different evidential questions.

The useful distinction is not “human memory bad, technology good”.

Clyde Street gives us too many counterexamples for that.

The useful distinction is that remembered observation needs to be treated as remembered observation.

A limit can sit in the image

Sometimes the observer’s method is not the main problem. The source itself is poor.

In July 2015, Keith reviewed the penalty shoot-out between Germany and France at the Women’s World Cup. The video he used was not of high quality, and he qualified the interpretation that followed: “From what I saw …”

That particular relationship between a visible limit and a constrained account is followed more fully in When a Limit Actually Constrains the Account. Here the observational point is narrower: the recording itself can bound what can be classified.

A better observer cannot recover detail that the recording does not contain clearly enough. Attention, memory, operational definition, observer drift and recording quality are not interchangeable failures, and they do not necessarily have the same remedy.

What looks missing may mean something else

That brings us back to the 2019 Asian Cup dataset.

Keith used vis_dat and vis_miss to inspect the data.

At first glance, the red-card and second-yellow variables were where the incompleteness appeared.

But Keith knew the semantics of his own coding.

NA meant that the card had not been awarded.

The “missingness” therefore had a different meaning from, say, a referee event that had happened but had not been recorded, or a value that the analyst did not know.

The software was not wrong to display NA.

The observer would be wrong to infer the wrong thing from it.

This is an important shift in the observation problem.

Nothing is blurry.

Nobody has forgotten the event.

There is no obvious observer drift.

The uncertainty sits in the relationship between representation and meaning.

The display shows a data state.

The interpretation depends on the codebook, convention or domain knowledge that tells us what that state signifies.

More sophisticated visualisation cannot remove the need for that knowledge.

It can make the need easier to notice.

Even an official source can change

Later in 2019, Keith found a different kind of boundary. The official World Rugby figures he was collecting could change after matches, particularly pass counts.

His response was to fix a twenty-four-hour capture rule for his own record. The custodial consequences of that decision — how another reader can know which state entered the analysis — are followed in What Has to Travel with Knowledge?

Here the observational point is narrower. “The official data” was not a timeless object. Two careful observers could consult the same official source at different times and encounter different values.

What was observed therefore depended partly on when the observation was made. Authority did not make the source static. That is not a problem a better pair of eyes can solve.

There is no single problem called observation

Across Clyde Street, observation keeps changing shape.

Sometimes the event is mediated by television.

Sometimes digital tools add layers of information that have to be filtered.

Sometimes the observer’s definitions, attention and reliability procedures matter.

Sometimes an eyewitness account belongs to memory rather than the live event.

Sometimes poor video constrains what can be classified.

Sometimes a missing value means “did not happen”, not “we do not know”.

Sometimes an official source changes after it has been observed.

It is tempting to treat all of these as versions of the same problem and reach for a single solution: better technology, more data, more cameras, more precise measurement.

Clyde Street does not support that confidence.

More technology can improve observation enormously.

It can also create more representations to interpret, more data to filter and more source states to track.

The right response depends on where the uncertainty lives.

If the recording is poor, the answer may be a better recording.

If the operational definition is vague, the answer may be a clearer definition.

If reliability has not been tested, another observer or repeated coding may tell us something useful.

If NA means “not awarded”, collecting another value would misunderstand the data structure.

If the official source changes, observations made at different times may not be equivalent.

And if the issue is the difference between real-time human arbitration and replay-mediated judgement, the question may be normative as well as technical: what kind of decision-making do we want the sport to have?

That is why the question “What can an observer know?” does not have one Clyde Street answer.

It has a prior question.

Question to carry

What kind of observation is this?

Only then can we ask what its limits allow us to claim.

Publication boundary note

This article owns the problem of differentiated observational knowability. It is not another general article about evidential limits, not a second video-literacy history and not an argument that technological observation is inferior to human judgement. Its central safeguard is that perception, memory, operational definition, observer reliability, recording quality, semantic missingness, representation and mutable source states are non-equivalent problems. Shared Clyde Street sources should link outward where another article already owns the deeper video, limitation, provenance or historical story.