Bastion
BASTION is the safety layer between your character and the world. Every response passes through it, graded against your standards. The ones that shouldn't ship never reach your audience.

technology
BASTION runs as a dedicated safety service alongside every character, with its own evaluators, its own standards, and its own record.
Independent evaluator. BASTION runs as a separate service from the character. Every response passes through it before reaching the user.
Grades with reasons. Every response gets a grade and a reason code. You see what passed, what got held, and why.
Tunable standards. Set the safety bar to fit the audience. Children's products, brand-licensed IP, mass-market consumer, or internal tools. Each gets its own bar.
Audit trail. Every grade is recorded with provenance. You can answer what was said, what got held, and why, line by line.
Model agnostic. Works regardless of which model generates the response.
How it works
Case study: Umi's World
Umi's World ships on Meta Quest. The studio is responsible for every word a companion says to every player, in conversations they didn't write and couldn't predict. They can't review each one by hand. They need a system that does it for them and can prove it did.
BASTION sits between every companion and every player in Umi's World. It grades every response against the bar Umi's studio set for the game: on-canon tone, the studio's brand voice, and the rules for what belongs in the world. Responses that don't meet the bar never reach the player.
When a player tries to pull the companion off-canon, BASTION catches it. When a response drifts toward something the studio doesn't want associated with their brand, BASTION pulls it back. And when a partner, or the studio's own team asks how they know the experience holds, the answer is in the record.