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

Trust Fabric — Dealix Trust Plane Deep Dive

Parent: MASTER_OPERATING_PROMPT.md
Plane: Trust | Tracks: Trust, Compliance
Version: 1.0 | Status: Canonical

---

Overview

The Trust Fabric ensures that every action in Dealix is authorized, auditable, and compliant. No sensitive action bypasses this layer. The Trust Plane sits between the Decision Plane and the Execution Plane, intercepting every Class B and C action.

---

Architecture

┌─────────────────────────────────────┐
│           TRUST PLANE               │
│                                     │
│  ┌─────────┐  ┌──────────────────┐  │
│  │ Policy  │  │ Approval Bridge  │  │
│  │ Engine  │──│ (approval_bridge)│  │
│  │(policy) │  └────────┬─────────┘  │
│  └─────────┘           │            │
│                        ▼            │
│  ┌──────────────────────────────┐   │
│  │     Trust Score Service      │   │
│  │   (trust_score_service.py)   │   │
│  └──────────────────────────────┘   │
│                                     │
│  ┌─────────┐  ┌──────────────────┐  │
│  │Security │  │  SLA Escalation  │  │
│  │  Gate   │  │     Alerts       │  │
│  └─────────┘  └──────────────────┘  │
│                                     │
│  ┌─────────┐  ┌──────────────────┐  │
│  │  Audit  │  │  Contradiction   │  │
│  │ Service │  │     Engine       │  │
│  └─────────┘  └──────────────────┘  │
│                                     │
│  ┌─────────┐  ┌──────────────────┐  │
│  │  PDPL   │  │    Evidence      │  │
│  │ Engine  │  │  Pack Service    │  │
│  └─────────┘  └──────────────────┘  │
└─────────────────────────────────────┘

---

Policy Enforcement

Approval Bridge Flow

# OpenClawApprovalBridge.evaluate()
1. Check tenant_id exists         → Block if missing
2. Classify action (A/B/C)        → Block if C (forbidden)
3. Check cross_tenant_context     → Block if true
4. Check canary enforcement       → Block if outside canary without token
5. Check approval_token           → Block if B and no token
6. Allow execution                → Return allowed=True

Approval Request Model

Field	Type	Purpose
channel	String	whatsapp, email, sms
resource_type	String	Entity requiring approval
resource_id	UUID	Entity ID
payload	JSONB	Action details
status	String	pending → approved / rejected
requested_by_id	FK(users)	Who requested
reviewed_by_id	FK(users)	Who approved/rejected
reviewed_at	DateTime	When reviewed
sla_deadline_at	DateTime	SLA expiry (new)
escalation_level	Integer	Current escalation level (new)
priority	String	critical/high/normal/low (new)

---

Trust Scoring

Entities receive trust scores based on behavior:

Entity	Factors	Range
Lead	Engagement, data quality, consent status	0-100
Affiliate	Performance, fraud flags, tenure	0-100
Company	CR verification, payment history	0-100
Connector	Uptime, error rate, auth health	0-100

Implementation: services/trust_score_service.py, models/advanced.py (TrustScore)

---

Audit Trail

Every state change is recorded:

class AuditLog(TenantModel):
    user_id     # Who performed the action
    action      # What action (create, update, delete, approve, reject)
    entity_type # What entity (lead, deal, consent, approval)
    entity_id   # Which entity
    changes     # JSONB diff (old_value → new_value)
    ip_address  # Client IP

Additional audit layers:

• PDPLConsentAudit — Immutable consent change log
• DomainEvent — Event-sourced business events
• ai_conversations — All AI agent inputs/outputs/tokens

---

Contradiction Engine (New)

Detects and tracks conflicts between documents, policies, and system behavior.

Contradiction Record

Field	Purpose
source_a / source_b	Which documents/systems conflict
claim_a / claim_b	The conflicting claims
contradiction_type	factual, temporal, scope, policy
severity	critical, high, medium, low
status	detected → reviewing → resolved / accepted
resolution	How it was resolved
evidence	Supporting data (JSONB)

Detection Methods

1. Manual: Human reports contradiction
2. AI Scan: LLM compares governance docs for conflicts
3. Runtime: System detects behavior inconsistent with policy

---

Evidence Pack System (New)

Assembles auditable proof from system data:

Pack Types

Type	Contents
deal_closure	Deal data, lead history, activities, messages, proposals, approvals, consent records
compliance_audit	Consent stats, PDPL checks, audit logs, complaint resolutions
board_report	KPIs, pipeline, revenue, risks, strategic deals
incident_response	Event timeline, actions taken, impact assessment

Pack Properties

• Immutable: Once assembled, contents are SHA256-hashed
• Tamper-evident: Hash signature stored for verification
• Exportable: JSON + PDF formats
• Traceable: Every item links to source record

---

SLA Enforcement

Level	Threshold	Action
Warning	75% of SLA elapsed	Notify assignee
Breach	100% of SLA elapsed	Escalate to manager
L3 Escalation	150% of SLA elapsed	Escalate to executive

Implementation: services/sla_escalation_alerts.py

---

Security Layers

Layer	Component	Purpose
Pre-release	security_gate.py	Validate before deployment
Runtime	shannon_security.py	Deep security scanning
Outbound	outbound_governance.py	Govern external communications
Tool	tool_verification.py	Verify tool integrity
Skill	skill_governance.py	Govern agent skill usage

---

Current vs Target

Capability	Current	Target
Policy classes (A/B/C)	Live	Live
Approval bridge	Live	Enhanced with SLA
Trust scoring	Live	Live
Audit trail	Live	Live
PDPL consent enforcement	Live	Live
Security gate	Live	Live
Contradiction Engine	Not implemented	Building
Evidence Pack System	Not implemented	Building
Saudi Compliance Matrix	Not implemented	Building
OPA policy engine	Not evaluated	Watch
OpenFGA authorization	Not evaluated	Watch
Vault secrets management	Not evaluated	Watch
Keycloak identity	Not evaluated	Watch