Eval checklist

A practical 30-minute matrix. Go in this order; stop at the first no-go.

Must-haves (15 minutes)

• Voice fingerprint + drift detection produces a usable score on your sample replies. Test on 10 of your last support transcripts.
• OpenAI-compat endpoint slot-fits where you currently call OpenAI (model field accepts persona ID).
• Free tier covers a 50-persona, 200-chat-session pilot.
• Data residency option matches your compliance posture (US or EU).
• Webhook delivery + signature verification is well-documented.

Should-haves (10 minutes)

• Provocation test suite passes ≥ 0.85 on a persona that represents your domain.
• Audit verdict aligns with your editorial judgment on 5 sample personas.
• SDK exists for your stack (TS, Python, Go, Ruby, Rust, Elixir).
• Pricing tier matches your projected volume.

Nice-to-haves (5 minutes)

• MCP server is available for your IDE.
• Marketplace listing UX is one you'd want your team to use.
• Lineage + fork tree maps onto your brand variant strategy.

No-go signals

If any of these are true, stop here:

• You need on-prem deployment (Moonborn is hosted only; ADR 0012).
• Your compliance regime requires HIPAA BAA but you're on a non- Enterprise tier.
• Your voice work needs Soul-level edits that contradict the four- layer model.
• You need a region that isn't US or EU (Asia Pacific not in v1).

"Later" signals

These aren't blockers but slow rollout — note them now:

• Marketplace commerce (paid listings) is Enterprise only — not blocking for a pilot, but blocks creator monetization.
• Custom moderation classifiers are Enterprise — start with defaults.
• More SDK languages — start with TS or Python; the others may lag for new APIs.

Output

Run this matrix and you'll have a Go / No-Go / Conditional in 30 minutes. Combine with the ROI calculation for the business case.

Related

• Vendor selection guide.
• Tier feature matrix.
• Pricing.