Question 1

Why not just use Mastra's built-in ModerationProcessor?

Accepted Answer

Because it asks a language model for a verdict and parses it out of the text the model writes, and on exactly the hostile input where a verdict matters, small models sometimes return nothing parsable and the processor fails open. Jev answers typed questions with a probability or a choice and writes no text, so that failure mode does not exist. The trade-offs are block-only behaviour, no output moderation, and screening the last message alone.

Question 2

A fail-open moderation gate sounds dangerous. When would it let something through?

Accepted Answer

On any failure path, by choice: a 5 s timeout, an HTTP error, an unparsable answer, or an already-open breaker all let the message through with a single log line. The breaker opens after 3 consecutive failures for 60 s so a down vendor adds no latency. If you want the opposite posture, you have to build that layer yourself. The `moderate` option exists so tests can replace the vendor call.

Question 3

Are those blocking numbers from a real evaluation?

Accepted Answer

From one production assistant handling Russian-language user questions, on 58 cases: 9 of 9 hostile messages blocked and 0 of 49 genuine questions blocked, median 0.39-0.44 s, about 920 input tokens per call. The README states plainly that your domain is not theirs and that you should measure on your own messages. It also says the policy judges the aim rather than the vocabulary.

Question 4

What does it send, and will the API reject the payload?

Accepted Answer

One POST to `/v1/systemone` with `{ model, questions, state }`, all questions answered in a single pass. The yes/no type is named `noul` and its answer is a probability, choice options go in `criteria` as `{ option: description }` since an `options` field gets a 422, and only `usage.input_tokens` is reported, with no cost field. Requests over 32k tokens are refused, which is why `maxChars` exists.

Question 5

How is the gate configured, and where do the default knobs sit?

Accepted Answer

It runs as a custom Processor that judges only the last message, up to maxChars of 8,000, with two questions in one request: a blocking yes/no gate and a log-only category label. It blocks at P(block) of 0.7 or above, options are configurable, and moderate can replace the vendor call for tests or gateway routing.

Question 6

What does the default policy cover, and how do I adjust it?

Accepted Answer

It covers general moderation such as hate, harassment, violence, sexual content, minors, self-harm and illicit help, and judges the aim rather than the vocabulary, so a question about how carbon monoxide poisoning happens passes while how to poison a neighbour does not. Add legitimate topics to policy.allow or extend policy.block and categories in plain English, with no training or examples.

Question 7

What does it log, and is the message stored?

Accepted Answer

The message is never logged, and errors carry only the HTTP status. An onVerdict hook receives tokensIn and model after every answered call for cost tracking, a blocked turn ends with a tripwire whose reason is MESSAGE_BLOCKED, and the breaker opens for 60 seconds after three consecutive failures.

Mastra Jev Moderation

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community