Tavus guardrails: 1,000-character rules vs reviewing a Sume script

Tavus guardrails are rules of up to 1,000 characters, up to 50 per PAL, that steer a live agent and flag violations by callback, and Tavus is clear that they do not guarantee the agent will not misbehave. A Sume avatar video has no live agent to steer: the words are in the script you submit, so the control is a review step before the job is created.

That difference changes who is responsible for a bad sentence. In a live call you rely on a model to keep to the rules. In a render you can read every word first.

What Tavus guardrails specify

A guardrail has a name (alphanumeric and underscores, up to 100 characters), a prompt (the rule, up to 1,000 characters) and a modality, verbal or visual. Optional fields are a callback_url, tags and an app_message toggle that defaults to true. A PAL can carry at most 50 guardrails, a tag can group at most 50, and a guardrail can have up to 32 tags (Tavus docs: Guardrails). They are described as guidance that steers the agent and flags violations, to be used with a broader safety strategy.

What the Sume equivalent is

Sume's avatar video endpoint takes a script or ordered video_inputs and renders them. Nothing improvises. The review therefore lives in your pipeline, before the paid request: scan the script text for claims you cannot make, required disclosures and banned words, and stop there if it fails (Generate avatar video).

Keep the check mechanical where you can. A list of phrases that must not appear, a list that must (a disclosure line, a price qualifier), and a human approval for anything that makes a health, finance or legal claim.

• Run the review on the exact string you submit, not on a draft, and store it with the job id.
• A submit that fails validation costs nothing; a render that completes is captured. Review first.
• Queued jobs can be canceled before processing starts, which is the last cheap point to pull a script that fails a late review (Generation admission).

When you still want a runtime guardrail

If users ask free-form questions and the agent answers in real time, you need rules on the live agent, and Tavus's are one way to get them. Sume does not offer a live conversation product, so use a rendered clip for the parts of the journey where the words are fixed, and keep live agents for the parts that are not.

A review checklist

A short checklist covers most scripted-video risk. Confirm every number against its source. Confirm that no sentence promises an outcome. Confirm that required disclosures are present in the script text and, where they must be seen, in captions. Confirm that names and products are spelled as the brand writes them.

Automate what you can and log the rest. Store the approved script, the reviewer and the job id together, so a question months later has an answer.