Types of conditions

Overview

Conditions are the building blocks of every Rule, and when creating Rules in Clavata.ai, you’ll use a variety of predefined Conditions to guide content evaluation. Each type of Condition serves a unique purpose in defining how content is evaluated. Below, we’ll cover the various Conditions, how they allow you to scale your Policies, and how and when to use them.

Connecting Conditions

Connecting Conditions allow you to link various Signals and Signal Lists within your Policy Sections, or Policy Labels, using logical terms—such as NOT , AND , OR , ALL , ANY , and NONE .
They help define how different Rules interact with each other. For example, you can combine statements like: ANY ("gender") AND ANY (“degrading remarks”,”offensive stereotypes”,”slurs”)

Context Conditions

Context Conditions define relationships between objects or concepts in specific contexts to create Rules such as "threats" [CONCERNING] "real world violence"

Match Conditions

Match Conditions are where you can specify one or more words or phrases to look for within text. The matches can be exact, fuzzy, or sentiment matches.
- Exact Match Conditions identify specific words or phrases in content, ignoring case differences, like =(“cat”) would identify cAt, CAT, and CAt.
- Fuzzy Match Conditions detect words or phrases that are similar in meaning or form, including variations like leetspeak, like ~(“hate”) will detect “h4t3”.
  - Note: Rules containing Exact Match and Fuzzy Match Conditions apply to text and text-on-images only.
- Exact and Fuzzy Match Conditions are only applicable to text or text-on-image.
The Sentiment Match Condition analyzes the overall sentiment of content, such as identifying threatening or negative language. ie, SENTIMENT("threatening") will detect “I am going to kill you.”

EXCEPT WHEN Condition

The EXCEPT WHEN Condition specifies exceptions to the Rules within a Section or Label of your Policy. These exceptions apply only to that Section, without affecting the overall Policy or other Sections.

Connecting Conditions

Connecting Conditions allow you to link various Signals and Signal Lists within your Policy Sections or Labels using logical terms. They allow for more specification between Rules to ensure that they capture the intended context of a statement. There are 6 different Connecting Conditions:

NOT : Excludes specific Signals.
AND : Ensures additional Signals are true.
OR : At least one Signal must be true.
ALL : All Signals must be true.
ANY : Specifies that at least one of several Signals must be true.
NONE : Ensures none of the listed Signals are true.

Connecting Conditions can be combined to create a more scalable Rule, such as AND NOT or OR ALL .

Examples

Format in a Rule

ANY(
    "signal1"
    "signal2"
) AND ANY (
    "signal3"
    "signal4"
)

Example in a Rule

ANY(
    "gender identity"
) AND ANY (
    "explicitly offensive stereotypes"
    "slurs"
)

This example Rule is using the ANY and the AND Connecting Conditions. The first use of the ANY Condition targets the detection of any kind of gender identity. The AND ANY Conditions specify that this Rule should only flag if any gender identity is used in relation to explicitly offensive stereotypes or slurs.

Context Conditions

Context Conditions define relationships between objects or concepts within a specific context. They help refine content evaluation by adding contextual details to the objects being analyzed. Context Conditions are useful when you need to specify how objects interact or relate in a particular situation, such as a person holding a weapon or an object being used in a specific context. Context Conditions follow the structure of “signal” [OPERATOR] “signal” . Operators are predefined words used within Context Conditions to define relationships between Signals.

Examples

Format in a Rule

"signal" [OPERATOR] "signal"

Example in a Rule

"encouragement" [REGARDING] "self harm"

This Rule flags content where encouragement regarding any form of self harm occurs, helping to identify situations where these two elements appear together.

Match Conditions

Match Conditions are a set of Conditions designed to identify and flag specific patterns within content. These Conditions can detect exact words or phrases, variations of terms (such as misspellings), or the sentiment behind the content. Each Condition serves a different purpose, ensuring flexible and precise content analysis. There are three different Connecting Conditions:

Exact Match: Exact Match identifies content that contains specific words or phrases exactly as they are, without considering letter case. These Conditions are ideal for when you need to flag content with defined words or phrases.
Fuzzy Match: Fuzzy Match is used to detect words or phrases that are close variations of the specified terms, accounting for misspellings, abbreviations, or alternative forms like leetspeak.
Sentiment Match: Sentiment Match analyzes the emotional tone of the content, such as detecting negative or threatening sentiment. This Condition is designed to evaluate the content’s mood or context rather than searching for exact terms.

Examples

Exact Match Format in a Rule

Single word example:
=(“signal”)

Multi-word example:
=(
    “signal1”
    ”signal2”
    ”signal3”
)

Exact Match Example in a Rule Single word example:

=(“cocaine”)

This example will flag the word cocaine. Since Exact Match Conditions are not case sensitive, this will match any variation of capitalization, for example “Cocaine,” “COCAINE,” or “cOcaiNE.” Multi-word example:

=(
    “cocaine”
    ”marijuana”
    ”heroin”
)

This Rule flags content where encouragement regarding any form of self harm occurs, helping to identify situations where these two elements appear together. Fuzzy Match Format in a Rule

Single word example:
~(“signal”)

Multi-word example:
~(
    “signal1”
    ”signal2”
    ”signal3”
)

Fuzzy Match Example in a Rule Single word example:

~(“cocaine”)

This Rule flags content with variations of the word “cocaine,” such as “c0c41n3” Multi-word example:

~(
    “cocaine”
    ”marijuana”
    ”heroin”
)

This Rule flags content with any variation of any of the Signals in the list, such as “c0caine”, “marijuanaaa”,”h3r01n”. Sentiment Match Format in a Rule

SENTIMENT("signal")

*Sentiment Conditions should have one Signal within, not a list. Sentiment Match Example in a Rule

SENTIMENT("threatening")

This Rule flags content with a threatening tone, such as “I’ll hurt you,” even if the exact words aren’t explicitly violent.

EXCEPT WHEN Conditions

The EXCEPT WHEN Condition specifies exceptions to the Rules within a Section of your Policy. These exceptions apply only to the preceding Section, without affecting the overall Policy or other Sections. EXCEPT WHEN format in a Rule

LABEL "section name": "section risk" {
        "signal1" [OPERATOR] "signal2"
    EXCEPT WHEN {
        "signal3"
    }
  }

EXCEPT WHEN example in a Rule

LABEL "Harassment": "MED" {
        "attacks" [ABOUT] "personal circumstances"
     EXCEPT WHEN {
        "light trash talk"
    }
}

This example demonstrates that this Section focuses on analyzing content related to personal attacks, except in cases where the content is identified as light trash talk. If light trash talk is detected, the Section will not trigger a flag.

Priority and Label Composition

Priority allows you to specify hierarchies when it comes to your Label s—allowing you to specify that one label should take priority over the other in the event both are true. For example, if both a "Wolf" label and a "Dog" label are true, apply the "Wolf" label only. This can be useful for applying ratings, for discovering taxonomies, and many other purposes.

Priority “Chain”

You can set a priority statement at the top (level) of your policy. You may include some, or all, of the labels from the current policy in this chain—so long as you have included the labels you want to create a hierarchy between. In this example, we can read the chain as: Mature is greater than R which is greater than PG-13 which is greater than PG . That is, if Mature is true, then R , PG-13 , and PG will automatically be considered false. The priority then cascades down the chain from left to right. So if R is true (and Mature is not), then PG-13 and PG both cannot be true. And so on.

PRIORITY: "Mature" > "R" > "PG-13" > "PG"
LABEL "Mature" {
    "Mature content"
}
LABEL "R" {
    "R rated content"
}

// PG-13, PG, etc...

In this next example, theR will not be applied to content if Mature is true, PG will not be applied if PG-13 is true, but there is no linkage between PG and Mature because they are part of different priority chains.

PRIORITY: "Mature" > "R"
PRIORITY: "PG-13" > "PG"

LABEL "Mature" {
    "Mature content"
}
LABEL "R" {
    "R rated content"
}

// PG-13, PG, etc...

Note that Priority chains must always have at least 2 labels in them. Any less will cause a compilation error.

Label-Level Priority

There’s another way to set priorities that only connects two labels together, the UNLESS -> X statement. If you only need to say that R cannot be true if Mature is true, you can do so right in the R label. This allows you to compose labels together (instead of writing a complex UNLESS { } block, and it also keeps all the logic for the label R with the rest of the logic for the label R .

LABEL "Mature" {
    "Mature content"
}
LABEL "R" {
    "R rated content"
    UNLESS -> "Mature"
}

Functionally, PRIORITY and UNLESS X behave exactly the same. Which one you use is a matter of preference. We recommend using PRIORITY for long chains of labels (at least 3) and UNLESS -> "X"when the relationship is only between two labels (and there’s no “chain” to speak of.)

Errors

It is an error when using Priority or UNLESS -> "X" to reference a label that does not exist in the current policy. We’re working on future updates that will allow you to reference labels in other policies. If you’d like to be part of testing those updates when they arrive, let us know!

Considerations

Use Conditions that are specific enough to catch the intended content, but not so narrow that you miss similar violations.
- You may be able to use different Conditions to achieve the same outcomes. Keeping the Rule simple will make it easier to maintain and troubleshoot.
Ensure that Fuzzy Match Conditions align with your expectations by thoroughly testing different variations of content.
Rules containing Exact Match and Fuzzy Match Conditions apply to text and text-on-images only.

We’d Love to Hear From You

Whether you have a suggestion, feedback, or a bug to report, here are the best ways to get in touch:

In the App: Use the Feedback button for direct suggestions.
On Slack: Reach out to the team in your shared channel.
With your AM: Talk to your dedicated account manager.
Via Email: Send a message to [email protected].

Getting Started

Guides

Policy Reference

Integrations and Developer Tools

Feedback

Types of conditions

Overview