compliance | /kəmˈplʌɪəns/ | noun the state or fact of according with or meeting rules or standards. all imports of timber are in compliance with regulations.
skimpliance | /skɪmˈplʌɪəns/ | noun the state or fact of appearing to meet rules or standards after briefly glancing at something. the mascot was in skimpliance with brand guidelines and in flagrant violation of every group chat in the country.
Starbucks Korea put a promotion live on 18 May 2026. The tumblers were shaped like tanks, so somebody called it Tank Day, and the tagline invited you to set your drink down with a satisfying tak. Everybody in the room thought this was just spiffing.
18 May is the day Koreans mark the Gwangju uprising, which the Chun Doo-hwan government put down by sending troops and tanks against pro-democracy demonstrators. The tak is worse. When the student Park Jong-chul died in police custody in 1987, the official story was that officers had smacked the desk and he had simply died of fright, and the phrase carrying that lie became a rallying cry for the democracy movement. Somebody had printed two national traumas from two separate decades onto a reusable cup, on the anniversary of one.
The CEO was dismissed with considerable briskness. The chairman apologised and afterwards sat through a lecture on modern Korean history, an education arriving a fortnight later than was perhaps entirely helpful. By August the business had gone from a comfortable operating profit to a loss of 18.4 billion won. In his apology the outgoing CEO said “the materials had not been thoroughly reviewed internally before the event began”, which is one way of putting it, and worth staying with, because what it describes happens constantly and is more interesting than carelessness.
The failures that matter are the ones where somebody checked
A review happens. Often several, at length, in writing, with a process wrapped round them. What the process reliably produces is paperwork, and the paperwork is wrong. The pattern sorts into species.
The handoff. On a Saturday morning in May 2013, an advert for the bookmaker Coral went out nine times between 06:42 and 08:42 on Cartoon Network and Boomerang. The ASA called it a serious breach. Coral said it was horrified, which one imagines it was. The ad had been cleared properly, and the clearance carried a note reminding broadcasters not to schedule gaming ads before nine at night. The restriction existed. It was written down. Then a booking number was allocated to the wrong advert and the whole thing sailed into Saturday morning cartoons. Compliance had done its job and produced a document, and the document lost a fight with a spreadsheet.
The wrong granularity. Channel 4 showed X-Men Origins: Wolverine at 18:55 on a Sunday in August 2012, with almost 12% of the audience aged four to fifteen. A film about a man whose defining professional skill is being aggressively knifey turned out to contain scenes unsuitable for children, to the astonishment of nobody except the people who scheduled it. Before you file this under nobody could be bothered, Channel 4 made twenty-seven separate edits to that film. Twenty-seven. Ofcom's finding came down to the observation that the problem had never been in the individual shots. It was the tone of the thing, which survives any amount of trimming.
Channel 4 also pointed out, reasonably, that a version of the film had aired pre-watershed before without complaint. Hold that up to the light. A version. The BBC holds 127 files of the same episode of Dad's Army, of which twenty are meaningfully distinct and the rest are duplicates in different formats and resolutions. One episode. One sitcom. The industry coined the word versionitis for this, which tells you how cheerfully everybody has made their peace with it. When somebody says the thing aired before and nobody minded, the follow-up is which one aired, and in most archives that cannot be answered inside a week. The filename is an assertion, as is the house ID, as is everything in the metadata, because all of it is what somebody typed at the time, in a hurry, possibly in 2004.
The unexamined claim. Persil ran a television ad describing itself as kinder to our planet, over footage of children clearing litter from beaches. The ASA banned it in 2022, having found the basis of the claim was never made clear and nobody had evidenced the product's full life cycle. Nothing in that ad looks dangerous. It is four words of warm sentiment on a bottle of detergent and it went through clearance without anybody's hair catching fire. Claims are the quietest liability in advertising, because whatever makes them a problem lives in a folder somewhere else, if it exists at all.
The case against all of this
Tank Day is obvious now. Whether it was obvious at four on a Thursday to a junior on a tumbler campaign is another question, and all of these look like negligence in retrospect while looking like an ordinary Tuesday at the time. Set them against the volume of advertising going out every year without incident and the failure rate approaches nothing. You could argue with a perfectly straight face that skimming allocates scarce attention sensibly, and that the occasional defenestration of a chief executive is what it costs.
What rescues the argument is that the knowledge nearly always sat inside the organisation already. Somebody at Turner knew that ad carried a nine o'clock restriction, because they had written it down. The information was in the building, just never in the room where the check happened.
Skimpliance is a rationing behaviour
Nobody sets out to skim. Skimming is what happens when checking is expensive and the queue is long, so deep review gets rationed by perceived risk. You check the regulated categories. You check whatever the client is nervous about. The tumbler campaign goes through on the nod, because it is a tumbler campaign and what could possibly happen.
Meanwhile the contexts multiply. Every asset goes out into territories with their own law and their own history, onto platforms with their own rules, into dayparts and feeds, on dates that mean different things in different places. AI has made producing those variants dramatically cheaper while making the reading of them no cheaper at all. The maw gets bigger. The number of people who know what happened in Gwangju in 1980 stays exactly where it was.
Which makes the useful question what would have to change for checking to stop being rationed. It turns out to be two information problems rather than an attitude problem.
You have to know what you have. Not what the metadata says you have. If one episode has 127 files behind it and twenty of them differ, every downstream check is running against a file you identified by hearsay. Fingerprinting content from its own image and audio turns which version is this into a lookup, and shrinks the queue, because you stop checking the same thing 127 times. Stripping out duplicates reads like a storage story. It is closer to the economic precondition for not rationing.
The same fingerprint answers the question that arrives the moment something goes wrong, which is where else is this. A problem clip is rarely confined to one ad. It's in the cutdown, the six-second version, the regional variant, the retrospective nobody remembers commissioning. Reverse clip search turns three weeks and a spreadsheet into a query.
You have to know what went out. Confirming what actually aired has conventionally taken around eight days in the UK. Eight days isn't a reporting lag, it's a window in which you can breach the same rule repeatedly before anybody troubles you with the news that it happened once. Coral aired nine times in two hours. Detecting playout in near real time closes that window to about a minute, which won't stop the first mistake but does rather limit the sequel.
The objection you have been loading
I have spent this whole article attacking organisations that produced evidence of checking in place of checking, and my answer produces evidence of checking.
Evidence becomes skimpliance when it stands in for the check, and it's fine when it falls out of one. The test is whether the artefact could have been produced without the check happening. A sign-off could. So could that scheduling note nobody read. A fingerprint couldn't, nor could a diff between an approved script and a delivered master, or a playout record taken off the broadcast signal. Those are the residue of a check rather than testimony that one happened, which is the through-line here. Every failure above involved trusting a description of the content over the content, whether that description was a filename, a booking reference, or a warm phrase about the planet nobody had evidenced.
What it won't fix
Tank Day would have got through.
Nothing that reads content against regulatory codes is going to know that 18 May carries weight in South Korea, or that a particular onomatopoeia has a 1987 death in custody attached to it. That knowledge is cultural and local and sitting in people's heads. A register of sensitive dates belongs somewhere queryable rather than in the memory of whoever happens to be in the meeting, but somebody has to know to write it down first. Two of the species above are now cheap to eliminate and one is reducible. The last stays a human problem, and clearing the others off the desk is what leaves attention available for it.
The remedy
Starbucks Korea has set out how it will stop this happening again. It will overhaul its review process by introducing checklists and artificial intelligence tools.
An organisation that produced evidence of review in place of review has resolved to fix the problem by producing rather better evidence of review. It might still work, and a checklist with does this date mean anything to anyone written on it would have saved them a chief executive and 18.4 billion won. Whether it becomes a check or remains an artefact depends on whether anybody can afford to treat it as one, which is a question about cost rather than conscience.
skimpliant | /skɪmˈplʌɪənt/ | adjective having been checked thoroughly, at the wrong level of detail. the film was fully skimpliant after twenty-seven separate edits.
Snicket Labs has spent seven years on the unglamorous half of this. Match identifies content from its own fingerprint rather than its metadata, reverse clip search finds every asset a given clip appears in, and Assure confirms what aired within a minute instead of eight days. The compliance layer between them starts landing in the Lab this quarter. If your compliance queue is a rationing exercise, we should talk.
