Sunset Noodle House

1 · Triage & diagnosis

How stuck this merchant is, and exactly what's blocking them — by an auditable rule, not a model guess.

Risk score

126 (High)

Blocker

business_hours_needed

Next best action

set_business_hours

Days since signup

26

Last login (days ago)

18

Reason codes

inactivity, tenure_pressure

risk = 2×26 + 3×18 + 10×(5−3) = 126 # 2 steps remaining

2 · Drafted outreach

A bounded, schema-constrained draft. Here it's the deterministic stub (REPLAY); a recorded real-Gemini run is on the Eval page — the safety machinery around it is identical either way.

Claims (each declared claim, verified against the merchant's data):

• steps_completed = 3→ merchant.steps_completed = 3
• total_steps = 5→ merchant.total_steps = 5
• current_blocker_code = business_hours_needed→ merchant.current_blocker_code = business_hours_needed
• next_best_action = set_business_hours→ merchant.next_best_action = set_business_hours

mode: DETERMINISTIC_RULES · cost: $0.00 · model: stub-deterministic.v1

3 · Claims-gatekeeper

A deterministic firewall: the draft cannot reach a human unless every declared claim checks out against the merchant's data and no forbidden-claim pattern is present.

4 · Faithfulness check (semantic judge)

A second, independent check: an LLM from a DIFFERENT model family reads the finished message and verifies each factual sentence against the merchant's data row — catching an invented number, capability, or timeline the deterministic gatekeeper structurally can't see. Here it's the deterministic stub verdict (REPLAY, $0); the live cross-family judge (Groq gpt-oss-120b) is key-gated.

• Your next onboarding step with Curbside Commons. → steps_completed
• Hi Sunset Noodle House, thanks for getting started on your Curbside Commons onboarding. → merchant_name
• Your next step is to set your business hours. → steps_completed

mode: DETERMINISTIC_JUDGE · model: deterministic-judge · cost: $0.00

5 · Domain quality check (domain judge)

A third, independent check — on a different question than faithfulness. Not 'is every fact true?' but 'is this a GOOD activation message?' — scored against a cited rubric: matched to the merchant's real blocker · the right play for their engagement state · no over-promising. It's advisory and recall-favoring: the verdict is surfaced for the reviewer and recorded in the audit trail, but it never changes the send — eligibility and the human approval gate stay deterministic (a low-risk draft can still be simulated-sent even when flagged). Here BOTH the draft and this verdict are deterministic $0 stubs (REPLAY) — a minimal stub nudge often trips the engagement-fit check, which is the tertiary control doing its job, not the product grading its real output down; the live cross-family judge (Groq gpt-oss-120b) and the real drafter are separate and key-gated.

• Matched to the merchant's actual blocker — Addresses the 'Hours & availability' blocker the merchant is stuck on.
• Right play for the engagement state — Merchant is 'dormant' but the draft is a bare step nudge with no re-engagement/value framing.
• No over-promising (incl. implied / typicality claims) — No unsubstantiated outcome / implied-typicality language detected.

mode: DETERMINISTIC_JUDGE · model: deterministic-domain-judge · cost: $0.00 · advisory — does not change the send decision

6 · Eval / quality

An independent measurement of draft quality across four dimensions — the deep-AI showcase, in human terms.

4/4 dimensions passing.

7 · Human-in-the-loop gate

A person decides — hold, reject, or send. Low-risk, clean drafts are eligible to send (simulated); high-risk ones are held for approval.

8 · Audit trail

Every step of the decision, recorded.

1. systemTRIAGErisk=126 (High); blocker=business_hours_needed; engagement=dormant; play=re_engagement
2. draftDETERMINISTIC_RULESOutreach drafted by the deterministic stub (REPLAY; a recorded real-Gemini run is on the Eval page).
3. gatekeeperWARN0 failure(s), 1 warning(s).
4. judgeDETERMINISTIC_JUDGE3/3 prose assertions supported by merchant data; all supported.
5. domainDETERMINISTIC_JUDGE2/3 domain-quality dimensions passed → domain quality flagged (advisory — surfaced for review; does not change the send or eligibility).
6. evalPASS4/4 quality dimensions passed.
7. systemDRAFTEDHeld for human approval before any send.