Untrusted text stays data
Email bodies, PDF text, page text and voice transcripts reach a model only inside delimited untrusted blocks, after hidden text has been pulled out as evidence of its own. A model never decides who receives mail: recipients, subjects and templates come from code.
Hard rules set the floor
When deterministic checks find critical evidence, such as changed bank details, a login form on a lookalike domain or a dangerous attachment type, the verdict cannot be Clean whatever a model says. A model can raise severity, never lower it below the floor.
Every claim cites evidence
Every sentence in a verdict must cite an Evidence ID from the checks that ran on that case. If a model cites a missing ID, the verdict is rejected, retried once, and then replaced by the rule-based verdict.
A sealed browser for links
When Muse Wonder walks a suspicious link, the driver, not the prompt, enforces the limits: http and https only, at most 8 steps and 60 seconds, no file downloads, and only per-case canary details typed into forms. Every action is logged.
Validated model outputs
Every model output and tool argument is checked against a strict schema before it is used, and tool calls run in auto mode only.
How we call Meta's Model API
Every call to Meta's Model API uses a Standard model ID with store: false, which asks Meta not to keep the request. Meta's API terms allow Meta to train on content only in Contributor mode, which we never use, and they publish no fixed retention period for requests.
Meta's Model API pricing and data terms