system-prompts-and-models-of-ai-tools/salesflow-saas/docs/governance/trust-closure-plan.md

Trust Fabric Closure Plan — Track 5

Parent: trust-fabric.md
Plane: Trust | Version: 1.0

---

Objective

Transform Trust Plane from "policy engine + audit logs" to "no sensitive action without approval + verification + evidence + correlation."

---

Live Trust Components Required

1. Approval Packet Flow (Priority 1)

Goal: At least one path where Class B action goes through structured ApprovalPacket → review → approve/reject → execute → evidence.

Target Path: WhatsApp outreach to new lead

Agent proposes send_whatsapp
  → ApprovalPacket schema generated (structured_outputs.py)
  → Policy gate classifies as B
  → ApprovalRequest created with SLA deadline
  → Reviewer gets notification
  → Approve → approval_token issued
  → OpenClaw gateway executes with token
  → Tool receipt generated
  → Evidence logged to ai_conversations + audit_log

Required Wiring:

• ApprovalPacket schema → approval_bridge.py integration
• SLA deadline field on ApprovalRequest model
• Notification to reviewer (email/WhatsApp)
• Evidence: approval_token + tool_receipt + audit_log linked by trace_id

2. Tool Verification Receipt Flow (Priority 1)

Goal: At least one tool call produces a verifiable receipt.

Implementation:

• tool_verification.py already exists
• tool_receipts.py already exists
• Need: receipts written for WhatsApp plugin calls
• Need: receipt includes trace_id, tenant_id, action, result_hash, timestamp

3. Contradiction Detection (Priority 2)

Goal: Real contradictions detected and flagged.

Implementation Plan:

• Wire contradiction_engine.py to CI pipeline
• On governance doc change: run LLM scan against other governance docs
• Store detected contradictions in contradictions table
• Show in Policy Violations Board frontend

4. Evidence Pack Viewer (Priority 2)

Goal: Unified evidence pack that links decision → tool → approval → output.

Implementation:

• evidence_pack_service.py exists
• Need: assemble_deal_pack that queries real data:
  • Deal from deals table
  • Lead from leads table
  • Activities from activities table
  • Messages from messages table
  • Approvals from approval_requests table
  • AI conversations from ai_conversations table
  • Consent from consents table

5. Trace Correlation (Priority 1)

Goal: trace_id / correlation_id links all related records.

Implementation:

• Add correlation_id to DomainEvent (already exists as field)
• Pass correlation_id through OpenClaw gateway → task router → agent → handler
• Store in ai_conversations.correlation_id, audit_log.correlation_id
• Query by correlation_id in evidence pack assembly

---

Watch Technologies — Adoption Criteria

OPA (Open Policy Agent)

Adopt when:

• Policy rules exceed 50 AND are complex (nested conditions, temporal logic)
• Current policy.py becomes maintenance burden
• ADR demonstrates value with prototype

Spike criteria:

• Prototype: 5 existing policy rules expressed in Rego
• Benchmark: latency comparison vs current Python implementation
• Integration: OPA sidecar evaluated for performance

OpenFGA

Adopt when:

• Authorization logic exceeds role-based (needs relationship-based)
• Multi-tenant permission inheritance becomes complex
• ADR demonstrates value with prototype

Spike criteria:

• Prototype: tenant → user → resource permission graph
• Benchmark: query latency for "can user X do action Y on resource Z"
• Integration: OpenFGA as authorization service evaluated

Vault

Adopt when:

• Secret rotation is needed for compliance
• 10+ distinct secret types managed
• Environment variables become unwieldy

Keycloak

Adopt when:

• SSO requirement from enterprise customer
• Multi-IdP federation needed
• Current JWT auth insufficient

---

Gate: Trust Closure

• One approval flow live end-to-end with SLA
• One tool verification receipt generated and stored
• One contradiction detected in real scan
• One evidence pack assembled from real deal data
• trace_id links decision → approval → execution → evidence
• Contradiction dashboard shows real data
• Approval SLA measured for at least one path