RFC-026: Human Feedback Signal — 2. Proposed Extension: FEEDBACK Message Type

AIGP Specification › RFC-026: Human Feedback Signal › 2. Proposed Extension: FEEDBACK Message Type

← 1. Problem Statement · Section index · 3. Evidence Storage →

2. Proposed Extension: FEEDBACK Message Type

2.1 New Message Type

{
  "protocol_version": "2.0",
  "message_type": "FEEDBACK",
  "app_id": "identity-service",
  "request_id": "req-1718500000000",
  "session_id": "sess-abc456",
  "pseudonym_id": "usr-a7f3b2c1e9d0",
  "timestamp": "2026-06-17T09:30:00Z",
  "rating": "thumbs_down",
  "comment": "",
  "nlp": {
    "sentiment": -0.6,
    "specificity": 0.85,
    "topics": ["iam", "cross-account"],
    "expectation_gap": true,
    "severity_language": "medium"
  },
  "context": {
    "user_query": "Review this inline policy for cross-account risks",
    "ai_response_snippet": "The policy grants sts:AssumeRole with Resource: * which..."
  },
  "relationship": {
    "type": "FEEDBACK_ON",
    "parent_id": "req-1718500000000",
    "parent_ref": "s3://evidence-bucket/.../req-1718500000000.json.gz"
  }
}

2.2 Field Definitions

Field	Type	Required	Description
message_type	string	Yes	Always "FEEDBACK"
app_id	string	Yes	Application that generated the AI response
request_id	string	Yes	ID of the invocation being rated
session_id	string	Yes	Session containing the invocation (RFC-010b)
pseudonym_id	string	Yes	Non-reversible user identifier (HMAC-derived)
timestamp	string	Yes	ISO 8601 UTC when feedback was given
rating	enum	Yes	"thumbs_up" or "thumbs_down"
comment	string	No	Optional user commentary (max 500 chars)
nlp	object	No	NLP analysis of comment (Tier 1 heuristics)
context	object	No	Snapshot of what was rated (for recall)
relationship	object	Yes	Parent-child linkage to the rated invocation

2.3 Pseudonym ID

The pseudonym_id enables behavioral analysis without identifying the user:

pseudonym_id = "usr-" + HMAC-SHA256(user_email, app_secret)[:12]

Properties:

• Stable: same user → same ID across sessions within the same app
• Non-reversible: governance server cannot resolve pseudonym → email
• App-scoped: different apps produce different pseudonym_ids (different secrets)
• Consent-compatible: user is never identified in the governance layer

2.4 NLP Analysis Object

Lightweight heuristic analysis of comment text (no LLM call required):

Field	Type	Description
sentiment	float	-1.0 (frustrated) to 1.0 (positive)
specificity	float	0.0 (vague) to 1.0 (precise technical detail)
topics	string[]	Extracted domain keywords
expectation_gap	boolean	True if “expected X but got Y” pattern detected
severity_language	enum	"low", "medium", "high" intensity

---

← 1. Problem Statement · Section index · 3. Evidence Storage →