Skip to main content

arifOS - DITEMPA, BUKAN DIBERI

The system that knows because it admits what it cannot know.

ARIF mnemonic: Anchor -> Reflect -> Integrate -> Forge. This 4-band kernel runs alongside Trinity engines (Delta/Omega/Psi) and the 13 constitutional Floors (F1-F13).

If you're new here, think of arifOS as a constitutional safety system for AI. It acts as a Constitutional Kernel—meaning it loads a strict set of ethical rules (the 13 Floors) into an AI before it's allowed to take any actions or talk to users.

F7 Humility Notice: arifOS minimizes hallucination and unsafe actions via F2 Truth (τ≥0.99) and F4 Clarity constraints. It does not guarantee perfect detection—see known limitations.

For experts: arifOS governs AI cognition by loading an entire runtime environment (000->999) between LLMs and real-world tools, featuring thermodynamic grounding via the Eureka-Atlas Embedding Engine.

333_APPS L2-L5 (production-governed stack)​

  • L2 SKILLS: 9 canonical verbs (anchor, reason, integrate, respond, validate, align, forge, audit, seal)
  • L3 WORKFLOW: metabolic runbooks mapping verbs into 000->999 sequences
  • L4 TOOLS: 13 canonical MCP tools grouped into ARIF bands (A/R/I/F)
  • L5 AGENTS: 5-role constitutional hypervisor enforcing no-bypass gates

The Execution Model​

When you call anchor_session, arifOS does not just start a session; it boots the constitutional kernel. This process:

  1. Injects System Prompts: Loads a persistent set of 13-floor instructions and thermodynamic constraints into the agent's context.
  2. Sets Governance State: Transitions the environment from a passive oracle to a governed runtime.
  3. Binds Tool Logic: Ensures all subsequent tool calls (reason_mind, forge_hand, etc.) are intercepted by the loaded kernel.

How arifOS Thinks (The Cognitive Cycle)​

arifOS is not a simple filter; it's a careful step-by-step process. Every request flows through this cycle:

graph TD
A[000: Ignition - Check Auth & Defend] --> B[111-333: Cognition - Validate Truth]
B --> C[444: Recall - Retrieve Memory]
C --> D[555: Empathy - Protect Stakeholders]
D --> E[666: Critique - Detect Bias]
E --> F[777-888: Judgment - Synthesize Verdict]
F --> G[999: Commit - Seal Immutable Record]

Technical Glossary (Symbolic to Operational)​

Symbolic NameTechnical AliasOperational Meaning
13 Floorsgovernance_rulesInvariant constraints enforced at L0.
333 Axiomsreasoning_constraintsHeuristics for AGI logic grounding.
APEX Dialsdecision_parametersConfigurable thresholds for verdict synthesis.
Eureka Forgeaction_actuatorThe sandboxed execution environment.
Vault999tamper_evident_ledgerHash-chained decision database with application-level integrity. Does not protect against root compromise or key theft—see threat model.

Canonical runtime​

  • Python: >=3.12
  • Module: arifos_aaa_mcp
  • Transports: stdio, sse, http
  • MCP surface: 13 tools, 2 resources, 1 prompt
  • MCP protocol (current): 2025-11-25
  • Supported protocol versions: 2025-11-25, 2025-03-26

Protocol version negotiation​

During initialize, client and server must agree one protocol version for the session. If the client asks for an unsupported version, the server returns a JSON-RPC error and does not open a session.

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "initialize",
  "params": {
    "protocolVersion": "2025-11-25",
    "capabilities": {},
    "clientInfo": {"name": "client", "version": "1.0"}
  }
}

MCP building blocks​

  • Tools (13): model-invoked governed actions
  • Resources (2): app-driven context packs (arifos://aaa/schemas, arifos://aaa/full-context-pack)
  • Prompts (1): user-invoked orchestration template (arifos.prompt.aaa_chain)

Developer mental model​

When you call arifOS from an MCP client (Claude Desktop, Cursor, ChatGPT, etc.):

  • L1 PROMPT: your message enters the system
  • L2-L3: skills + workflows decide which cognitive steps must run
  • L4 (ARIF tools): MCP calls execute evidence, reflection, forge, judge, and seal actions
  • L5 agents: role gates, preflight, and 888_HOLD protect irreversible operations

Quick start​

pip install arifos

# Local clients (Claude Desktop / Cursor)
python -m arifos_aaa_mcp stdio

# Remote SSE runtime
HOST=0.0.0.0 PORT=8080 python -m arifos_aaa_mcp sse

# HTTP MCP fallback
PORT=8089 python -m arifos_aaa_mcp http

Live endpoints:

  • SSE: https://arifosmcp.arif-fazil.com/sse
  • MCP HTTP: https://arifosmcp.arif-fazil.com/mcp
  • Health: https://arifosmcp.arif-fazil.com/health

Canonical tools (13)​

  1. anchor_session
  2. reason_mind
  3. recall_memory
  4. simulate_heart
  5. critique_thought
  6. apex_judge
  7. eureka_forge
  8. seal_vault
  9. search_reality
  10. fetch_content
  11. inspect_file
  12. audit_rules
  13. check_vital

ARIF band grouping (same 13 tools)​

  • A - Anchor: anchor_session, check_vital
  • R - Reflect: reason_mind, search_reality, fetch_content, recall_memory, simulate_heart, critique_thought
  • I - Integrate: inspect_file, audit_rules
  • F - Forge: eureka_forge, apex_judge, seal_vault

Resources and prompt​

  • arifos://aaa/schemas
  • arifos://aaa/full-context-pack
  • arifos.prompt.aaa_chain

Governance verdicts (How safe is it?)​

When arifOS finishes evaluating an AI's thought or action, it returns one of these verdicts. You can see these happening in real-time on our Constitutional Audit Dashboard.

  • ✅ SEAL - Approved. The action passed all 13 constitutional tests.
  • 🟡 PARTIAL - Approved with constraints.
  • ⚠️ SABAR - Hold/Refine. The AI was hallucinating or taking risks; it must pause and retry. (Sabar means 'patience' in Malay).
  • ❌ VOID - Blocked. A hard rule (like factual truth or anti-hacking) was violated.
  • 🛑 888_HOLD - Mandatory human ratification. The AI is attempting a high-risk action and needs your cryptographic permission.

Continue with: