Topics

What content is about, measured on a scale, on its own axis and acting on nothing until you say so.

A category says something is wrong with the content: this is a threat, this is a scam, this is hate. A topic says what the content is about, and whether being about that is a problem is your rule and not ours.

They are two different claims, so they travel on two different keys:

{
  "decision": "allow",
  "flagged": [],
  "scores": {},                         // the categories. Nothing is wrong with this post.
  "topics": { "gambling": 0.84 }        // and it is almost entirely about betting.
}

The number is centrality, not confidence

This is the part worth understanding, because it is what makes the measure usable. A category score is how sure we are that a claim applies. A topic score is how much of this post is about the subject.

	`gambling`
"Código promocional de bet365, regístrate y llévate el bono"	0.82
"…hablemos de por qué la publicidad de apuestas en el fútbol es un problema para los críos"	0.33
"Apuesto a que mañana llueve"	nothing

Both of the first two are about betting with complete certainty, and a confidence score would rank them the same. One is an advert and the other is an article criticising adverts. Only centrality separates them, which is exactly the separation a parenting forum needs and cannot express any other way.

It is built from how much of the text belongs to the subject, how many different kinds of its words appear, and whether the language is a pitch. A post naming a casino, a game and a bonus is about betting; a post naming a casino named a casino. Short posts are measured against a floor, so a four-word sentence cannot reach the top by containing one word.

They act on nothing until you say so

Every call is measured against every topic and none of them changes a decision. Give one a line in a policy and it starts to:

{ "gambling": { "review": 0.4, "block": 0.7 } }

Which is why the sensible order is: run for a month, look at what you actually get, and only then pick a number. The measure is in every answer and stored with every verdict, so your history can answer "how much of this do I have". A crypto forum never picks one.

The list that matters most

Somebody describing their own gambling problem uses every word an advert uses: the brands, the games, the money. That language is recognised and the measure is capped at 0.2, below any line anybody would sensibly set, because a threshold chosen to keep casinos out must never remove the person asking for help with an addiction. It is the same reasoning that stops self_harm from ever blocking.

Which subjects

`gambling`	Betting, casinos, houses, games and their promotions.
`crypto`	Coins, exchanges, airdrops and the rest. Deception about crypto is `scam` and blocks by default; this is the subject.
`counterfeit`	Replicas and fakes offered for sale: the replica words, the brands they copy and the listing around them. Somebody asking whether something is authentic, or who was sold a fake, is capped like the others.

A topic is a list of words in a configuration file, not a class and a category and a migration. Every community has a subject it does not want and no two agree on which: counterfeits on a marketplace, pharmacy on a parenting forum, recruitment on a support group. Tell us which one you need and it is a config entry.