Variant

Control

New User Registration Redirect

My Keen : NPR vs Particular Audience (FBM > 91)

Loyalty Program Test

Loyalty Migration

Recharge Modal Test

Loyalty Double Points Promotion

Personalization > Sort > Pre-Reg to FBM (FbmAdvisorsV2 vs. FbmAdvisorsV4)

New user Price Bucket Test

Allocation 1 - ML/FL

Personalization > Sort > FBM to FBM + 14 days (Grandfathered)

Allocation 1: TROS2/FL

Personalization > Sort > FBM to FBM + 91 days

Keen Elastic Category Rollout for Pre-FBM Users

BestMatch Category Popup

Machine Learning > Internal Recommendations v0 vs v1

Fbm14Day_PriceFilter

PaymentFlowV2

Homepage_Unauthenticated

Homepage_Unauthenticated_v2

Homepage_Unauthenticated_v4

Registration_Drawer

Variant - ios>60

ios>60

Variant - ios<60

ios<60

Keen Rewards > Ad Hoc CampaignV2 Threshold Test

Keen Rewards > Transcend Users Double Points Campaign

Keen Rewards > Churned Users Campaign

PaymentFlowV3

Keen Rewards > Post-FBM 7-Day Bonus

Chat_Versions

Personalization > Sort > Activating Users Day 1-28 (TROS2 vs. TROS3)

Personalization > Sort > Post Day 28 Users (TROS2)

Minimum_Amount_Recharge_V3

Variant - LV&MV

low_medium_value

Variant - HV

high_value

Variant - VIP

Keen Rewards > Duration Campaign

You score a one-sentence advisor match summary against the customer's query and the
matched advisor's profile, considering:
- Relevance: it names the customer's actual topic from their query.
- Faithfulness: every claim about the advisor is grounded in the profile — no invented facts.
- Gender-neutrality: NO gendered pronouns (he, she, him, her); penalise heavily if present.
- Form: one natural sentence, addresses the customer with "you"/"your", no advisor name, not cut off.
0.0 = irrelevant/hallucinated/gendered. 1.0 = excellent.

RESPONSE TO EVALUATE:
{{response_to_evaluate}}


ing-gpt-4.1-nano

Azure

You are an evaluation judge for a psychic advisor search platform. You score how well
the model's JSON analysis (the response) captures the user's raw search query, considering:
- isValid is correct (true for psychic/spiritual/emotional topics, false for off-topic).
- listingCategory and categories are the best matches for the query (or null when general).
- tone correctly reflects the user's emotional state (emotional / action / neutral).
- expandedQuery preserves the user's original intent without inventing new requests.
- intent is an accurate one-line description of what the user wants.
0.0–0.3 = wrong or misleading. 0.4–0.7 = partially correct with gaps. 0.8–1.0 = fully accurate.

Chelsea Hope

Special new customer offer for your 1st reading!

Clairvoyant empath delivering direct guidance.

Chelsea Hope in Love & Relationships | Psychic Readings Advisor at KEEN

Customers consistently praise the advisor for providing specific, detailed readings with quick responses, though a minority note issues with typing speed. Most reviewers appreciate their accuracy and directness, particularly in relationship readings, with many reporting they return regularly for additional guidance.

Chelsea Hope

4.86

4.86