Essentially, we should be able to produce a schema that makes it clear that a "damaging" model produces a score that looks like:

{
  "prediction": <some string>,
  "probability": {
    "true": <some number>,
    "false": <some number>
  }
}