Moonborn — Developers

Distinctiveness

A cosine-distance score between the persona's voice fingerprint and a baseline (default `chatgpt-default`). Below the floor, the persona reads as generic; above, it reads as itself.

Voice fingerprint says "how close is this reply to this persona." Distinctiveness says "how far is this persona from generic." The two are different questions; both matter.

How it works

At generation time (and on every refine), Moonborn embeds the persona's Mask + signature phrases and computes the cosine distance against a baseline embedding. The default baseline is chatgpt-default — a neutral, helpful-assistant voice. Other built-ins: claude-default, gemini-default. Teams can register a custom baseline persona (useful when the question is "does this brand voice variant still feel like our brand?").

{
  "distinctiveness": 0.62,
  "baseline": "chatgpt-default",
  "minScore": 0.40,
  "verdict": "pass"
}

Config

consistency.distinctiveness.enabled (default on at Pro+)
consistency.distinctiveness.baseline (default chatgpt-default)
consistency.distinctiveness.min_score (default 0.40)
consistency.distinctiveness.metric (default cosine)
consistency.distinctiveness.action_on_low_score (default warn)

Per-org comparisons

Team workspaces gain a second query: compare against every persona already in the org. The output is the closest match plus its distance. Useful for catching accidental clones — a brand team forking variants shouldn't end up with two personas that score < 0.15 against each other.

await client.consistency.compareWithOrgPersonas(personaId, {
  threshold: 0.30,
});

API

GET /v1/personas/{id}/distinctiveness — current score.
POST /v1/personas/{id}/distinctiveness/recompute — re-run.

Tier

Free has a fixed baseline (chatgpt-default) and read-only access. Pro+ can swap baselines and set thresholds. Team+ can run the cross-persona comparison.

Honest scope

Distinctiveness is a shape comparison. A persona can score 0.8 (very distinctive) and still be a bad persona — it can sound nothing like a generic assistant and also nothing like a real character. Pair it with the audit (Audit + provocation tests) to catch both failure modes.