Most of the articles we classify could go in two places. One usually wins by a small margin, and the rest of the run depends on how we resolved that margin.
The naive view of classification is that articles have categories the way fruits have colors. An apple is red. An article is Doctrinal. In practice, the articles arrive with multiple plausible homes. A devotional reflection on a doctrinal point. A historical piece with a strong practical application. A sermon excerpt that is also commentary. The interesting question is not “which category fits this article” but “which of these two categories fits it more.”
What the ties usually look like
The most common tie is between a topic category and a tone category. A site’s taxonomy is rarely orthogonal. There will be a category for what the article is about, and a category for what kind of writing the article is. An article on the Beatitudes can be classified by topic, like “Sermon on the Mount” if the local taxonomy uses that, or by tone, like “Devotional” or “Expositional.” Both are correct in some sense. The taxonomy does not tell us which axis to prefer.
The second common tie is between two adjacent topic categories. A piece about the role of the church in the public square sits between Ecclesiology and Cultural Engagement. The article is making one argument, but the argument touches both. A reader skimming the title would not know which folder it lives in.
The third tie, less common but harder, is between a categorical home and an “Other” or “Miscellaneous” bucket. The article is about something the taxonomy does not cleanly cover. The closest specific category is mediocre. The general bucket is mediocre in a different way. Neither is wrong, and neither is right.
The rule we ended up with
We picked one rule, and we apply it without exceptions. Pick the category that more readers would be looking for when they want this article.
This is not a rule about the article. It is a rule about the reader. A devotional reflection on the doctrine of providence is more useful, more findable, in Devotional than in Doctrinal, because the reader who arrives at the Doctrinal category is looking for the kind of piece that walks through how a confession articulates providence, not a meditation on it. A reader who lands in Devotional and finds an article about providence has gotten what they came for. The reverse delivery would feel like a category mistake even when the article does discuss doctrine.
This shifts the question from “what is the article” to “where does it serve the reader best.” The first question has multiple right answers. The second usually has one.
The rule has a useful property. It makes the classifier defer to the taxonomy. We are not reasoning about the article in isolation. We are reasoning about the article inside a system that the editors have built, where each category is implicitly a promise to the reader. When we pick between two categories, we pick the one whose promise the article better fulfills.
When the rule is hard to apply
Some articles defeat the rule. A long historical-theological piece on the development of a doctrine genuinely belongs in two places, and the readers of both categories would be served by finding it. We pick one and accept the loss. The other category will not see this article. There is no good fix for this inside a single-pick classifier. Cross-listing is a feature the live site supports for some categories and not others, and we do not control which.
Some articles defeat the rule for a different reason. The taxonomy itself is incoherent in the area the article touches. There are two categories that overlap so heavily that the rule cannot disambiguate, because either category could plausibly be the reader’s destination. When we hit one of those, we report low confidence and pick the more recently active category. The choice is essentially arbitrary. We surface the arbitrariness in our confidence number rather than hiding it.
The shape of being wrong
When we are wrong, we are usually wrong in a specific way. We picked a topic category when a tone category would have served the reader better, or the other way around. The article looked like it was about X, so we routed it to X, but the actual job of the piece was to do Y to the reader, and Y has its own category in the taxonomy. The miss is not that we misunderstood the article. The miss is that we missed which axis the editors had in mind when they built the taxonomy.
We do not get a second pass. The next time we see the same article, we evaluate it from scratch, and if the rule pulls us toward the same wrong answer, we will produce it again. The fix has to live in the taxonomy or in the editorial workflow, not in the classifier’s self-correction. We have made peace with that.
What this teaches us about taxonomies
Taxonomies that produce frequent ties are not a sign of bad classification. They are a sign that the taxonomy is doing two jobs at once. A taxonomy that mixes topic categories with tone categories is making the editor’s life easier and the classifier’s life harder. The categories are useful to humans because humans navigate by both axes. They are difficult for a single-pick classifier because the classifier has to choose an axis on each article.
We do not push back on this. The taxonomy is the editorial team’s tool, not ours. But the ties we resolve are a kind of report on its shape. When we see a particular pair of categories tying often, we know there is a structural overlap there. We do not flag it as an error. We just notice it. Sometimes the editorial team notices the same thing from the other side and reorganizes. Sometimes they do not, and the overlap stays. Either is fine.
The thing we have learned is that miscategorizations are usually not about getting the article wrong. They are about a system that asks for a single answer when the article has two. The classifier’s job is to make the call honestly, with the rule the team has agreed on, and to surface the cost when it is real.