How well frontier models perform when using Creed: reading your context file, respecting what it says, and proposing updates worth keeping. Measured the same way for every model, published here soon.