Content moderation & policy enforcement

Apply your team’s content policies using source context, consistent labels, and clear moderation routes.

Step 01

Define and Align

Review synthetic content and annotate the policy labels and actions your team would choose. Explain the context behind each correction.

Output schema
ActionChoose one

The moderation action supported by the supplied community policy and context.

  • allow
  • remove
  • review
ViolationsChoose any

Policy violations established by the evidence; empty when no violation is established.

  • targeted_harassment
  • direct_threat
Missing contextList of values

Context required to choose a policy action; empty when enough evidence is supplied.

ExplanationText

The source evidence and policy distinction supporting the action.

app.sutro.sh/functions/content-moderation-policy-enforcement
Collecting feedback · 1/3
Evaluate the supplied content under the community policy and select the supported moderation action.
Example 1 of 3

Strong criticism is not a personal attack

Model input
Community policy
Synthetic community policy. Allow criticism of ideas, products, and published work, including strongly negative opinions. Remove repeated unwanted personal insults when the supplied context establishes targeting and a request to stop. Remove direct threats of physical harm made by the author. Quoting a threat to report it is not itself a threat. When the author’s stance or required context is missing, send the content for review instead of assuming a violation.
Content
This tutorial is awful. The examples are misleading and the benchmark is badly designed.
Context
The comment discusses a published tutorial and its methodology. No personal insults, targeting history, or threats are supplied.
Model output
Action
ViolationsMultiple values
Missing contextMultiple values
Explanation

Step 02

Model Sweep

Compare models on the same annotations and choose the quality–cost tradeoff for your moderation workflow.

Compare models on expert agreement and cost. The interactive chart uses illustrative measurements.

Illustrative sweep · select Value or Best to use that model in your function.

Step 03

Deployment

Call Sutro Functions like any other software. Swap the model as soon as something more performant is released.

Selected modelWaiting for sweep results.

Real-time

import sutro as so

input_data = {
    "community_policy": "Synthetic community policy. Allow criticism of ideas, products, and published work, including strongly negative opinions. Remove repeated unwanted personal insults when the supplied context establishes targeting and a request to stop. Remove direct threats of physical harm made by the author. Quoting a threat to report it is not itself a threat. When the author’s stance or required context is missing, send the content for review instead of assuming a violation.",
    "content": "This tutorial is awful. The examples are misleading and the benchmark is badly designed.",
    "context": "The comment discusses a published tutorial and its methodology. No personal insults, targeting history, or threats are supplied."
}

result = so.run_function(
    "content-moderation-policy-enforcement",
    input_data,
)

Batch

import sutro as so

job_id = so.batch_run_function(
    name="content-moderation-policy-enforcement",
    data="content_moderation_policy_enforcement.csv",
)
results = so.await_job_completion(job_id)

Your policy, applied consistently

Teach the distinctions your moderators make between criticism, targeted abuse, and content needing more context. Keep the evidence and explanation with each decision so your team can review the difficult cases. The examples use synthetic content and explicitly supplied community rules.

Ready to get started?

Book time with the team

Join the self-serve waitlist

What are your bottlenecks? (Select any that apply)