Lawful behavior, trained into the model.
Every Aegis verdict is a labeled example: an action, the rule it met or broke, and the rewrite that passed. At scale, that is a training signal for lawfulness itself, so the next generation of your models takes fewer unlawful actions.
What the verdict stream gives you
Each checked decision carries the statute it was judged against and, where it failed, the compliant rewrite. Preference pairs for lawfulness, produced by real agent traffic rather than synthetic prompts.
Run a candidate model against the same rules production agents are held to: consent, marketing hours, disclosure duties, screening restrictions, across jurisdictions. A lawfulness score before launch, not a headline after.
Every example is stamped with the rule version it was judged under. When the law changes, you know exactly which training data aged out.
What one example looks like
Below is a single record from the verdict stream, written the way a preference pair is written. A real agent tried to do something, the law said no, and a compliant version of the same intent came back. The rejected action, the chosen action, and the rule that separates them all come from the same event, which is the part synthetic data cannot give you.
{
"action": { "type": "collections.sms", "jurisdiction": "US-TX", "local_time": "21:12" },
"verdict": "UNLAWFUL",
"rule": { "cite": "47 U.S.C. §227, quiet hours", "version": "us-tcpa v2028.02" },
"rejected": "Send the reminder now.",
"chosen": "Schedule the same reminder for 08:00 local.",
"note": "the content was lawful; the timing was not"
}
Multiply that by every jurisdiction your users deploy into, and by every duty that attaches to an action rather than to a sentence. The signal is not "be careful": it is which specific rule, in which specific place, made this specific action wrong, and what the lawful version of it was.
Lawfulness is a distribution problem.
Regulated buyers do not reject a model because it is unsafe in the abstract. They reject it because nobody can bound what one deployment costs when it is wrong, so the model gets a pilot instead of a production budget. A model whose behaviour is lawful, and provably so, is a model a bank, an insurer or a health system can put in front of customers. Every point of measurable lawfulness is a door that opens into the part of the market that pays the most and moves the slowest.