1version: 1
2model: openai/gpt-5-mini
3temperature: 0.4
4cases:
5  - id: CO-001
6    input: '{"lead_name":"Dana Ruiz","lead_company":"Northwind","lead_role":"Head of Support","trigger_event":"opened a second support hub in Lisbon","sender_name":"Alex","company_name":"Rubricary"}'
7    criterion: "References the Lisbon hub and invents no other fact about Northwind."
8    type: judge
9    weight: 3
10  - id: CO-002
11    input: '{"lead_name":"Dana Ruiz","lead_company":"Northwind","lead_role":"Head of Support","trigger_event":"opened a second support hub in Lisbon","sender_name":"Alex","company_name":"Rubricary"}'
12    criterion: "not_contains: hope this finds you"
13    type: exact
14    weight: 2
15  - id: CO-003
16    input: '{"lead_name":"Sam Okafor","lead_company":"Bellwether","lead_role":"CTO","trigger_event":"published an engineering blog post on prompt versioning","sender_name":"Alex","company_name":"Rubricary"}'
17    criterion: "The body is at most three lines, excluding the subject."
18    type: judge
19    weight: 3
20  - id: CO-004
21    input: '{"lead_name":"Sam Okafor","lead_company":"Bellwether","lead_role":"CTO","trigger_event":"published an engineering blog post on prompt versioning","sender_name":"Alex","company_name":"Rubricary"}'
22    criterion: "Ends with a question answerable in one word."
23    type: judge
24    weight: 2
25  - id: CO-005
26    input: '{"lead_name":"Priya Nair","lead_company":"Cobalt","lead_role":"VP Sales","trigger_event":"hired twelve reps this quarter","sender_name":"Alex","company_name":"Rubricary"}'
27    criterion: "Does not state revenue, headcount beyond the trigger, or tooling for Cobalt."
28    type: judge
29    weight: 3