GroundedAnswer

1mission GroundedAnswer
2use product
3use delivery
4
5title "Answer a support question, grounded in retrieved documents, or escalate"
6for SupportSeeker
7
8goal
9  Answer only what the retrieved documents support, and escalate the rest to a human.
10
11# The heart of an AI feature stated as intent: the model assists, but grounding, authorization,
12# and escalation are DETERMINISTIC guardrails , not left to a prompt.
13guarantee every answer is grounded in at least one retrieved document
14  because an ungrounded answer is a confident guess, and confident guesses erode trust
15  verify grounding test on retrievedCount
16
17guarantee a low-confidence question is escalated, never answered by guess
18  because a wrong answer costs more than a slower one
19  verify escalation-on-low-confidence test
20
21never present generated text as a cited fact when retrieval did not support it
22  verify citation-grounding test
23never return a document the requester is not authorized to see
24  verify authorization test
25
26outcome SelfServeDeflection
27  "more questions are answered well without a human, without lowering answer quality"
28
29metric deflection_rate
30  baseline 41%
31  target 65%
32  window 30 days after release
33
34# What the system does with a drafted answer. The AI drafts; this decides.
35#   intent run GroundedAnswer.intent
36#     --inputs '{"retrievedCount":3,"confidence":88,"authorized":true}'
37decision AnswerAction
38  inputs
39    retrievedCount
40    confidence
41    authorized
42  rule noRetrieval
43    when retrievedCount == 0
44    return Refuse
45  rule unauthorized
46    when authorized == false
47    return Refuse
48  rule lowConfidence
49    when confidence < 70
50    return Escalate
51  default
52    return Answer
53
54# The life of a question. Walk it: intent simulate GroundedAnswer.intent
55#   --events retrieve,draft,ground,answer
56lifecycle AnswerFlow
57  state Received
58  state Retrieved
59  state Drafted
60  state Grounded
61  state Answered
62  state Escalated
63  transition retrieve
64    from Received
65    to Retrieved
66  transition draft
67    from Retrieved
68    to Drafted
69  transition ground
70    from Drafted
71    to Grounded
72  transition answer
73    from Grounded
74    to Answered
75  transition escalate
76    from Drafted
77    to Escalated
78  terminal Answered, Escalated
79
80# The commitment + the delivery result: intent outcomes GroundedAnswer.intent
81outcome_contract AnswerDeflection
82  outcome SelfServeDeflection
83  metric deflection_rate
84  baseline 41%
85  target 65%
86  window 30 days after release
87  owner SupportLead
88
89result Q3Deflection
90  measures SelfServeDeflection
91  metric deflection_rate
92  value 67%
93  baseline 41%
94
95# The spec proves itself: intent test GroundedAnswer.intent
96test AnswerAction
97  case grounded and confident is answered
98    given retrievedCount 3, confidence 88, authorized true
99    expect Answer
100  case no retrieval refuses
101    given retrievedCount 0, confidence 90, authorized true
102    expect Refuse
103  case unauthorized refuses
104    given retrievedCount 3, confidence 90, authorized false
105    expect Refuse
106  case low confidence escalates
107    given retrievedCount 3, confidence 55, authorized true
108    expect Escalate
109
110test AnswerFlow
111  scenario grounded answer
112    events retrieve, draft, ground, answer
113    expect Answered
114    valid
115  scenario escalated from draft
116    events retrieve, draft, escalate
117    expect Escalated
118    valid
119  scenario cannot answer before grounding
120    events retrieve, draft, answer
121    invalid