The Mislabel in the Esports Data Pool: An Azur Lane Cosplay Set and the Cost of Careless Classification

On Tuesday evening, I was running a content inventory sheet for a client in Chicago — a company tracking esports media, wanting to know by what percentage the volume of esports articles had risen over the same period. I scrolled down and stopped at a line tagged "esports." The headline was about a cosplay photo set of the character Shimakaze from Azur Lane. I opened it, read it from start to finish, then read it again because I thought I had missed something. No tournament was mentioned. No team, no professional player, no coach, no balance patch, no transfer contract, no standings. The only thing present was a photo set: a person in a sailor outfit with a pair of rabbit ears, posing against a plain backdrop, accompanied by a few lines praising the performer for a fairly impressive recreation of the in-game original. I noted it in my spreadsheet. Every number is a story waiting to be verified, and this article, as a data row in the sample I was responsible for measuring, was a number. A number placed in the wrong slot. That was the start of a week I spent answering a seemingly small question: how did a cosplay photo set get filed by an automated classification system into the same drawer as transfer news, match schedules, and statistical analysis? Azur Lane is a mobile gacha game developed by Manjuu and Yongshi. It first launched in China in May 2026, reached Japan in September of that year, and came to the international market in May 2026. It is not a competitive title with a professional tournament system. It has no regional qualifiers, no franchise slots, no world championship. Its operating rhythm is entirely different: every few weeks, the publisher releases a new character or a new outfit, and the community responds with fan art, short videos, and cosplay. Shimakaze is one of the game's most recognizable characters — a destroyer of the Sakura Empire in the game's setting. The original article described her as hard to overlook among a crowded roster, with an instantly recognizable design and the ability to transform through many outfits. That is the language of character marketing. Not the language of game balance. The article was published on a Vietnamese-language outlet, credited to an author named Tuấn Hưng. The notable part sits in the link block below: it pointed to a PUBG Asia Stars event, a dispute involving a PUBG player, and a headline about a gaming media company director wanted for copyright infringement. This outlet exists by mixing content types — cosplay photos, genuine esports news, copyright news within the gaming industry — and that is a completely legitimate traffic strategy from a business standpoint. The problem lies elsewhere. When an outlet mixes genres, the data collection system cannot read context. It reads tags, keywords, and links. That is where everything begins to break apart. Before drawing conclusions, I do what I always do: reconstruct the definition. Content is classified as esports when it contains at least one of the following — an organized competition; a professional team or player; a transfer event within a league framework; a game update affecting competitive balance; or a publisher governance decision relating to competitive integrity. Five criteria, independent of one another. I checked the article against each criterion. Organized competition: no. Professional team or player: no. Transfer event: no. Balance patch: no. Governance decision: no. Five out of five criteria returned negative. The only thing the article contained was two entities: a content creator who produced the photo set, and a character within a game. There was no third entity that could be called a competitive subject. Data never lies, but the people who define it can. In this case, no one lied. There was only a labeling system working off surface signals. The article mentioned a video game. It sat next to articles mentioning PUBG — a genuine esports title. The URL, the keywords, and the links all coexisted in one cluster. The classifier saw the co-occurrence and applied the tag. That is the fundamental flaw of every frequency-based classification system: it measures lexical proximity, not substantive proximity. Those are entirely different things. Two articles both about "games" can belong to two different economies, serve two different audiences, and operate on two different cycles. And this is where I want to linger longer, because simply saying "mislabeled" keeps the story shallow. The economy this article actually belongs to is the fan-content economy built around a character IP. I call it the gacha IP flywheel. It works like this. A publisher designs a highly recognizable character — easy to draw, easy to remember, easy to cosplay, with visual elements distinctive enough to stand out on a phone screen. That character is released alongside a limited outfit. The community responds with fan art, short videos, and cosplay. Each of those content layers pulls more people into the game. The game gains revenue. The publisher gains budget to design the next character. The loop closes. The cycle of this flywheel is not a patch cycle. It is a banner cycle. There is no buff, no nerf, no tier list. There is a new character, a new outfit, and a wave of accompanying fan content. The performer of the photo set appears here as a traffic node, not a competitive subject. The article praised their presentation in aesthetic language — a mischievous spirit, closeness to the in-game original — rather than in performance language. No engagement metrics were published. No follower count, no engagement rate, no share count. The conclusion that the photo set was "carefully invested" is a promotional claim, not a measurement. As an analyst, I do not diminish that economy. It is real, it is large, and it is measurable. But it is not measured with esports instruments. The wrong metric is more dangerous than measuring nothing at all. If I used match viewership to measure the heat of a cosplay set, I would conclude the cosplay set failed. If I used cosplay engagement to measure the popularity of a tournament, I would conclude the tournament is booming. Both conclusions would be meaningless. Now comes the quantification, because this is the part that bothers me most. Suppose a client pipeline collects one thousand articles tagged esports in a month. If the noise rate from cosplay and fan content sits at four percent — an entirely reasonable figure for outlets that mix genres — then forty articles in my sample contain no competitive subject whatsoever. Those forty articles lift the total content volume index, but contribute nothing to any analysis of tournaments, teams, or players. They dilute the sample. They make an upward-trending chart look as though esports is growing, when in fact only cosplay is growing. If that rate reaches seven percent — and during weeks with major character events it can — the noise figure is seventy articles. At that scale, you are no longer measuring esports. You are measuring an unseparated mixture. I did not realize this in a meeting room. I realized it while doing exactly what I do every week: reviewing the article list to flag cases needing manual review. Based on my years of experience tracking matches and content streams, I had developed a simple reflex: before filing content into the competitive category, I look for a scoreboard. If there is no scoreboard, no schedule, no roster, then that content does not belong to the competitive category, regardless of what the headline says. That reflex sounds obvious. But it is not obvious to an automated pipeline. A pipeline has no reflex. A pipeline only has labels. This is the point I want to stress to anyone building a content data system: label error is not a small bug to be fixed by hand. It is a systemic fault. Every mislabeled article flows into every subsequent report, every subsequent model, every subsequent decision. It does not vanish on its own. It accumulates. And it accumulates in a specific direction: inflation. Because the "esports" label typically carries higher commercial value than "cosplay" or "fan content," in any system where labels are applied ambiguously, the tendency leans toward the more attractive label. This is a form of inflation. No one actively creates it. It simply happens. But I must argue against myself, because that is my rule. Is it fair to blame the outlet? I do not think so. That outlet is doing its job: publishing content its readers want to read. The cosplay set has an audience. The PUBG news has an audience. Blending the two is a reasonable editorial decision for a general-interest outlet. The problem is not created by the writer. The problem is created by the data consumer — the systems, including mine, designed to count volume rather than count substance. We build machines that reward article counts, then act surprised when article counts rise by every means possible, including means unrelated to the subject we intend to measure. There is a correlation here easily mistaken for causation. Both esports and cosplay sit within a broad space called "gaming culture." Because they share that space, people easily assume they share a measurement system, an audience, a growth logic. They share none of those. A gacha game can have tens of millions of players and not a single professional tournament. An esports title can have millions of tournament viewers, most of whom have never played that game for more than ten hours. The same "game" label on two entirely different objects. I once made a similar mistake in another form. In 2026, I published an expected-goals model for a World Cup group stage and wrongly concluded that a national team should have won. A veteran analyst pointed out that I had failed to subtract the shot-angle coefficient and defender pressure, inflating the metric by more than thirty percent. I spent six weeks reviewing all sixty-four matches and recalibrating the model. The lesson was not that the model was wrong. The lesson was that I had drawn a conclusion before understanding the definition of my own tool. The error in today's problem belongs to the same family. Not bad data. But an unverified definition. So what signals should be tracked in the next cycle? Separate content verticals before measuring anything. If you are measuring esports, define esports clearly through testable criteria — scoreboards, rosters, schedules, contracts, regulations. Any content that fails that test should sit in its own vertical, no matter how attractive its headline. Track the noise rate as a standing metric. Not a one-time check and then forgotten. It fluctuates by season, by event, by character release cycle. If you do not track it, it will silently change the meaning of every other number. There is a genuine esports story sitting right beside this article. The link block pointed to a PUBG Asia Stars event and a dispute involving the player Himass, along with a headline about a gaming media company director wanted for copyright infringement. Those are pieces of content with real competitive subjects and real governance actors, and they deserve separate analysis with suitable tools. They do not belong in this article, but they belong in my tracking book. The last thing I took from this week has nothing to do with Azur Lane, nothing to do with cosplay, and nothing to do with any specific game. It has to do with the habit of checking definitions before trusting numbers. Every match is a data sample, but belief is the only variable that cannot be entered into a spreadsheet. And when I look at a list of one thousand rows of content tagged esports, what I am really checking is not which game is trending. What I am checking is whether my system is seeing the same thing I am seeing. This week, it was not. That is why I had to rewrite the definition.