🛸 Agentic Intelligence Governance Manual (V2)

Status: ISO-parity strategic reference
Architecture: n8n ➜ LangGraph ➜ OpenClaw
Goal: Turn operational know-how into secured technological autonomy.

Manual scope
This in-app manual is an operational introduction.
It explicitly distinguishes: In place today, Rolling out, and Target.
It is not a full technical inventory of all application surfaces.

Executive promise (one page)

• Built-in confidentiality: sensitive data is protected before processing.
• Permanent control: humans stay in charge, including emergency stop.
• Managed profitability: AI power matches task value.
• Business continuity: the firm stays operational in degraded mode.

📑 Structure & navigation

Pillar	Operational focus	Master tool	Emojis
I. Doctrine	Decision cycle	n8n ➜ LangGraph ➜ OpenClaw	🧠 ⚙️
II. Presence	Action & learning	LangGraph	📝 📈 🚦
III. Convergence	Multi-agent coordination	LangGraph	🛰️ ⚖️ ⚓
IV. Emergence	Strategy & calibration	Knowledge graph	🔬 🧪 🌐
V. Security	Infrastructure & cage	Security lock	🛡️ 🏗️ 🗝️ 👁️ ♻️
VI. Efficiency	Performance & costs	D1–D4 matrix	🧩 🔋 ⚡ 🔒

🧠 I. Doctrine: the core decision cycle

The system rests on an architectural invariant that strictly separates perception, reasoning, and action.

⚙️ Technical articulation READ — THINK — WRITE

[ Perception ] [ Reasoning ] [ Action ]
   (READ)       (THINK)       (WRITE)
      |             |             |
[n8n] --------> [LangGraph] --------> [OpenClaw]
      |             |             |
 Incoming       Confront with     Secured
 signals            rules         execution

• READ — ⚙️
  • Role: Captures inbound signals (email, webhooks, APIs).
  • Action: Immediate wash: clean and anonymize data before analysis to protect trade secrets.
• THINK — 🧠
  • Role: The “brain”. It confronts signals with your rules.
  • Action: Analyzes context, weighs risk, and prepares an intent to act.
• WRITE — ⚙️
  • Role: The “arm”. It physically executes the task.
  • Action: Fires only after brain validation and lock release.

Usage: two operating modes

• Cockpit mode (daily): in Presence , you consume decisions, supervise intents, and arbitrate quickly (5-minute daily ritual).
• Lab mode (strategic): in Emergence , you evolve business laws when context changes (about 1 hour/month or on regulatory change).
• Delegated supervision: simple approvals (D1/D2) can be delegated to team leads; leadership keeps sensitive approvals (D3/D4) and law editing ownership.

The economic framing of D1–D4 is detailed in Efficiency . The printable due-diligence evidence pack is consolidated in Strategic FAQ .

Dieidos — Pilot intelligence, delegate action. 🛸🔒

🔒 Parity lexicon (business ↔ technical)

Summary
Each business term maps to its technical concept and architectural role, as in manual V2.

Parity lexicon (business ↔ technical)

Business term	Technical concept	Architectural role
Perception	READ	Capture and sanitization ⚙️.
Reasoning	THINK	Brain and analysis 🧠.
Action	WRITE	Execution in an isolated environment ⚙️.
Law / rule	Rule	Business instruction carved in stone.
Maturity	Threshold	Promotion threshold toward autonomy.
Draft	Draft	Intent to act, pending approval.
The lock	Guardrail	Physical execution blocker.
The cage	Node Host	Secured, ephemeral execution space.

❓ VII. Compliance Pack (executive format)

Summary
This page is a printable due-diligence evidence pack: human governance, control, traceability, and continuity.

Due-diligence evidence (defensible)

• Effective human supervision: AI proposes and prepares; approvals follow explicit governance.
• Controlled delegation: D1/D2 approvals can be delegated to mission leads; D3/D4 approvals remain under leadership control.
• Emergency control: if behavior drifts, the stop mechanism can immediately freeze automated actions.
• Usable traceability: every significant decision leaves a reconstructible trace for audit or dispute.

What if AI stops? (simplified BCP)

• Business continuity: the firm keeps access to its data and business tools.
• Reversibility: assisted mode can be slowed or frozen without corrupting case data.
• Controlled recovery: restart keeps decision history and the same governance frame.

Executive questions (short answers)

1. Is it easy to change a business rule? Yes. Daily supervision happens in Presence; law editing happens in Emergence to publish durable adjustments.

2. Who approves what in the organization? Simple approvals (D1/D2) can be delegated; leadership keeps sensitive approvals (D3/D4) and law governance.

3. What happens if AI attempts an unauthorized action? The system blocks execution, raises an alert, and returns to human control through escalation procedure.

🗂️ VERSIONING: documentation history

This page describes the product history of documentation inside the application. It is separate from the V2 governance manual (business content), whose reference text lives in the repo at ops/docs/Manuel de Gouvernance Agentique Complet.md.

Active version

• V2 — Agentic intelligence governance manual (web integration FR + EN)
• Languages — French + English
• Source — same manual as end-user documentation (ops file above)

Change log

1. Migration of the manual to dedicated MDX pages.
2. Added trust guarantees and visual proofs.
3. Added strategic Mermaid diagrams.
4. Added navigation and documentation search.

Alignment note
This content reflects a consolidated product target and is updated as delivery iterations evolve.

📝 II. PRESENCE: the daily cockpit

Summary
Presence is the supervision surface: read the situation, choose posture, certify a technical (WRITE) action, then verify the actual execution outcome. Detailed law and business-label editing happens in Emergence.

Cockpit mode: daily usage

• Cadence: about 5 minutes per day.
• Goal: consume reliable decisions, not perform engineering.
• Expected outcome: fast arbitration and operational continuity.

Presence keeps decisions on one screen:

• Situation and rules — idea, active rule, modality gauges (visible steering evidence).
• Perspective and perimeter (side panel) — the “why”, then the “how far”, including the verb applied to the selected rule.
• Integrated validation — no standalone pending panel; validation runs in the current cockpit flow on the active rule. Approve (reinforce) dispatches execution and anchors the decision for this pass.
• WRITE feedback in Perspective — the right panel shows the current WRITE state (success, failed, pending) and reflects it in Perspective history per agent/idea context.
• Reject and correct — opens Emergence with agent, idea, rule, and action reference prefilled, plus the reason you entered, so you can adjust the law without losing context.

Supervision delegation

• Typical firm setup: D1/D2 actions can be approved by mission leads in the Presence flow.
• Leadership governance: D3/D4 approvals remain under leadership control.
• Role separation: operational validation happens here; law editing happens in Emergence .

The 5-minute ritual and D1–D4 are covered under Efficiency — performance and costs . Printable due-diligence supervision evidence is summarized in Strategic FAQ .

🛰️ III. CONVERGENCE: the team command post

Summary
Convergence keeps multiple LangGraph reasoning streams coherent: coordination, friction arbitration, strategic anchoring.

UI availability
Convergence remains a core product pillar. Its direct exposure in navigation can vary by rollout context.

Convergence manages coherence across several LangGraph reasoning flows.

• Multi-agent coordination
  • Analysis: Oversight of the team (legal, sales, support agents).
  • Link: Ensures one agent’s WRITE ⚙️ actions do not contradict another’s THINK 🧠 analyses.
• ⚖️ Friction arbitration
  • Signal: “Friction” appears when two rules collide.
  • Lanes: Visualization of the different choices available to the system.
  • Decision: You act as arbiter to set the final priority.
• ⚓ Strategic anchoring
  • Action: Change a major priority in one click.
  • Impact: Instantly realigns all agents to a new “world view”.

🔬 IV. EMERGENCE: rules as firm capital

Summary
Emergence is the primary place to calibrate laws: business labels (ontology overrides), idea chain, directive, and simulation before publish. A correction started from Presence lands here with context (agent, idea, rule, action reference, reason).

Lab mode: strategic usage

• Cadence: about 1 hour per month, or when regulation changes.
• Goal: evolve firm rigor, not process daily flow.
• Expected outcome: stronger and transferable business rules.

Emergence complements the Presence cockpit:

• Contextual editing — drafted rule, verbs, linked ideas, directive; labels shown in Presence follow these overrides (canonical ids remain for audit).
• Simulation — controlled space to test behavior before real execution.
• Handoff from Presence — explicit banner and reusable reason when you correct after a rejection.

Intangible firm asset

• Durable capitalization: published business laws remain in the engine even as teams change.
• Know-how transfer: onboarding accelerates because operating rules are already formalized.
• Firm ownership: you document for controlled execution, not for shelfware.

Guarantee

The draft ➜ simulation ➜ publish path stays auditable; humans remain the final authority on which laws apply.

⚙️ READ — Perception & sanitization

Summary
READ captures inbound signals then applies an immediate wash. Sensitive data is neutralized before any analysis.

READ — ⚙️

• Role: Captures inbound signals (email, webhooks, APIs).
• Action: Immediate wash: clean and anonymize data before analysis to protect trade secrets.

Role in the cycle

READ is the perception step of the manual’s cycle (Section I — Doctrine): it feeds THINK without exposing raw identity. For downstream execution (lock, cage), see Security — trust infrastructure .

🧠 THINK — Reasoning & rules

Summary
THINK is the “brain”: it confronts signals with your rules and prepares an intent to act, without executing a physical action alone.

THINK — 🧠

• Role: The “brain”. It confronts signals with your rules.
• Action: Analyzes context, weighs risk, and prepares an intent to act.

Guarantee

THINK stays separate from WRITE: reasoning alone never authorizes a physical action.

⚙️ WRITE — Secured action

Summary
WRITE is the “arm”: physical execution of the task, only after brain validation and lock release.

WRITE — ⚙️

• Role: The “arm”. It physically executes the task.
• Action: Fires only after brain validation and lock release.

Guarantee

The system applies prior certification before WRITE and prefers safe blocking when in doubt.

🛡️ V. SECURITY: trust infrastructure

Summary
Security separates reasoning (LangGraph) from execution (OpenClaw): physical lock, cage, vault, red button, GDPR purge.

Security physically separates reasoning (LangGraph) from execution (OpenClaw). Upstream washing and neutralization are described under READ — perception and sanitization .

• Physical security lock
  • Role: Unbreachable gatekeeper to the real world.
  • Function: Physically blocks any signal to OpenClaw until proof of authorization is provided.
• 🏗️ The cage
  • Concept: Every WRITE ⚙️ action runs in a sealed space (Node Host).
  • Security: The execution environment is created for the task then destroyed immediately. Zero data leakage between cases.
• 🗝️ Vault & confidentiality
  • Secrets: The agent uses your access (mail, CRM) via the Vault without ever being able to read or memorize credentials.
  • Anonymization: The AI reasons on “neutralized” data (blurred KPIs) to protect trade secrets.
• 🚨 Red button (break-glass)
  • Role: Absolute takeover.
  • Action: An emergency mechanism can instantly freeze all AI actions to return to 100% human control.
• ♻️ Automatic purge (GDPR)
  • Compliance: The system applies operational data purge mechanisms according to retention policies.
  • Guarantee: The platform is designed to respect the right to be forgotten and legal retention cycles.

Continuity and reversibility

• Business continuity: if AI is paused, the firm keeps access to operational data and business tools.
• Reversibility: assisted mode can be suspended without damaging active casework.
• Recovery: restart follows the same safety and governance rules.

🧩 VI. EFFICIENCY: performance & costs

Summary
D1–D4 matrix, fuel strategy, and the 5-minute ritual in Presence for controlled cognitive ROI.

The daily ritual lives in Presence — daily action cockpit .

The four reasoning levels (D1–D4)

      /  D4  \      <-- Strategy & simulation (expert)
     /   D3   \     <-- Authoring & precision (powerful)
    /    D2    \    <-- Drafts & daily work (standard)
   /     D1     \   <-- Triage & filtering (flash)

Level	Task type	Engine used
D1 / D2	Triage & daily drafts	⚡ Fast & economical
D3 / D4	Expertise & simulation	🧠 Powerful & precise

• 🔋 Fuel strategy
  • Optimization: The system automatically picks the least costly AI for each cycle step.
  • Cognitive ROI: You pay maximum power only for strategic decisions.
• ⚡ The 5-minute ritual
  • Method: 5 minutes per day in Presence to validate and teach 📈.
  • Result: A machine that automates itself 🚦 day by day.

Usage governance (organization)

• D1/D2: approvals can be delegated to mission leads to absorb daily flow.
• D3/D4: sensitive approvals stay with leadership.
• Emergence: law editing remains restricted to authorized roles.

Economic degraded mode

• Under budget pressure, the system prioritizes D1/D2 usage instead of abruptly stopping operations.
• High-value tasks can be rescheduled under leadership validation.
• The firm keeps operational continuity with an explicit cost-control frame.

👁️ Sovereignty: audit & transparency

Operational complement to manual V2: audit chain, evidence in incidents, and takeover. The manual’s “Security” pillar covers lock, cage, and break-glass; this page formalizes continuous traceability.

Summary
Every important decision leaves a trace. In an incident you can freeze, analyze, and recover. Governance is demonstrable, not only declarative.

Active governance

• Log key decisions.
• Prove the validation path.
• Guarantee immediate takeover (break-glass).

Defensible evidence (leadership, regulator, insurer)

• Demonstrable diligence: human supervision is explicit and documented.
• Readable decision chain: who approved what, and in which context.
• Defensible traceability: in audit or dispute, each decision can be explained without black-box opacity.

Guarantee

Transparency does not rely on a verbal promise. It relies on an explicit chain of events, validation, and audit.