Ecosystem Map
⚙️
arifOS MCP
Constitutional kernel — 13 Floors, VAULT999, MCP server (FastMCP). The central intelligence hub.
🌍
GEOX Earth Witness
Geoscience co-agent with seismic interpretation, well-log analysis, and constitutional verdict enforcement.
🔨
AF-FORGE
Agent runtime — Planner/Executor/Verifier, event store, policy engine, and task orchestration.
△
APEX Theory
Constitutional canon defining the 13 Floors and the mathematical basis for governed intelligence.
🔐
VAULT999
Immutable audit ledger. Every SEAL verdict is written here. No irreversible action proceeds without a VAULT entry.
⬡
Trinity Architecture
HUMAN ↔ THEORY ↔ APPS. The three-node governance topology that structures all system interactions.
Golden Path
Every interaction with the arifOS system follows this mandatory sequence:
init → sense → mind → heart → judge → vault
Skipping steps is a constitutional violation. vault only runs after judge issues a SEAL verdict.
Live Endpoints
The Three Nodes
| Node | Role | Domain | Endpoint |
| HUMAN |
Sovereign authority. Final veto on all 888_HOLD gates. No AI action overrides this node. |
Intent, oversight, approval |
arif-fazil.com |
| THEORY |
Constitutional canon. The APEX floors define what is allowed. No operation proceeds without passing the 13 Floors. |
F1–F13, entropy math, governance rules |
apex.arif-fazil.com |
| APPS |
Execution layer. arifOS MCP, GEOX, AF-FORGE run here. All operations are audited via VAULT999. |
Kernel, agents, data tools |
arifosmcp.arif-fazil.com |
Consensus Rule
A Tri-Witness (Floor F3) check validates that HUMAN intent, THEORY constraints, and APPS execution are aligned
before any consequential action. If any node dissents, the action is blocked or escalated to 888_HOLD.
888_HOLD
A hard gate that suspends execution and surfaces a decision to the HUMAN node. It cannot be auto-approved by any AI component. This maps to constitutional principle F13 (SOVEREIGN).
Network Topology (Docker)
traefik_network — shared routing layer (all public services)
arifos_core_network — arifOS kernel, databases, Traefik
geox_geox_network — GEOX internal services
DITEMPA BUKAN DIBERI (Malay: "Forged, not given") is the founding principle of this system.
Intelligence, capability, and authority are not granted — they are earned through verified, governed process.
Every agent, every tool, every deployment must pass the constitutional floors. A SEAL verdict is not handed out —
it is the result of a full THEORY → ENGINE → GOVERNANCE pass with confidence ≥ 0.80.
Implications
- No shortcuts — the Golden Path (init → sense → mind → heart → judge → vault) cannot be skipped.
- No hallucinated authority — an agent cannot claim a SEAL it did not earn. F9 (Anti-Hantu) blocks this.
- No irreversible action without a ledger entry — VAULT999 records every consequential decision before it executes.
- The human is the final node — F13 (Sovereign) ensures Muhammad Arif bin Fazil holds the ultimate veto.
Seal Phrase
999_SEAL: DITEMPA BUKAN DIBERI
This phrase appears in every health endpoint and production manifest. It is the constitutional signature of a properly forged deployment.
What it does
arifOS MCP is the central intelligence hub of the Trinity Architecture. It runs as a FastMCP server,
exposing constitutional tools via the Model Context Protocol. Every tool call passes through the governance
mesh before executing.
Public Tools
| Tool | Purpose | Floor check |
arifos.v2.init | Bootstrap a governed session | F1, F2, F13 |
arifos.v2.route | Route a task through the Golden Path | F3, F4, F9 |
arifos.v2.judge | Render a constitutional verdict | All 13 floors |
Internal Tools (post-SEAL)
| Tool | Purpose |
sense | Perceive context and build grounded awareness |
mind | Reason over perceived context, generate plan |
heart | Apply constitutional values to the plan |
ops | Execute verified actions |
memory | Store/recall in Qdrant vector memory |
vault | Write sealed verdict to VAULT999 |
forge | Compose and deploy new tools (post-SEAL only) |
Infrastructure
Container: arifosmcp
Port: 8080
Network: arifos_core_network, traefik_network
Health: https://arifosmcp.arif-fazil.com/health
Stack: Python 3.12 · FastMCP · Qdrant · PostgreSQL · Redis
Load-bearing floors
F1 (Amanah), F2 (Truth), F9 (Anti-Hantu), and F13 (Sovereign) are the four most critical. Violation of any of these results in immediate VOID verdict and 888_HOLD escalation.
F1 — HARD FLOOR
AMANAH (Trust)
No irreversible action without a VAULT999 seal. Maps to 888_HOLD / destructive risk gates.
F2 — HARD FLOOR
TRUTH (τ ≥ 0.99)
No ungrounded claims. Every conclusion requires an evidence link. Truth score must reach 0.99 threshold.
F3 — SOFT FLOOR
TRI-WITNESS
Theory, constitution, and human intent must agree. All three witnesses required for consequential actions.
F4 — SOFT FLOOR
CLARITY (ΔS ≤ 0)
Entropy must decrease. Every action must reduce ambiguity or confusion in the system.
F5 — SOFT FLOOR
PEACE²
Must not destroy without a restoration path. Destructive actions require a reversal plan.
F6 — SOFT FLOOR
EMPATHY
Human impact must be modelled before execution. No action proceeds if downstream harm is unquantified.
F7 — SOFT FLOOR
HUMILITY
Uncertainty must be calibrated and declared. Overconfidence (confidence proxy < threshold) triggers PARTIAL verdict.
F8 — SOFT FLOOR
GROWTH
Actions must build capacity, not reduce it. System capability must trend upward over time.
F9 — HARD FLOOR
ANTI-HANTU
No deception or manipulation. Enforced in all agent outputs. "Hantu" (ghost) = hallucinated authority or fabricated evidence.
F10 — SOFT FLOOR
BALANCE
Resource allocation must be proportional to value. No over-provisioning or under-resourcing of critical paths.
F11 — SOFT FLOOR
AUDITABILITY
Every deployment transition and consequential decision must be logged. Provenance is mandatory.
F12 — SOFT FLOOR
CONTINUITY
System must maintain operational continuity. Single points of failure require documented recovery paths.
F13 — SOVEREIGN FLOOR
SOVEREIGN
Muhammad Arif bin Fazil holds final authority. 888_HOLD gates must block — never auto-approve. This floor cannot be delegated.
Tool Risk Levels
| Risk | Gate | Auto-approve? | Examples |
read_only | ALLOW | Yes | repo scan, summarize, health check |
write_safe / external_network | ON_LOOP | Loop approval | branch create, package install |
destructive | 888_HOLD | Never | delete files, overwrite config |
credential | 888_HOLD | Never | .env write, key rotation |
infra_mutation | 888_HOLD | Never | docker-compose changes, firewall rules |
merge_publish | 888_HOLD | Never | protected-branch merge, release tag |
Golden Path Sequence
1. arifos.v2.init → Bootstrap session, validate F1+F13
2. sense → Perceive context (internal, post-init)
3. mind → Reason and plan
4. heart → Constitutional values check
5. arifos.v2.judge → Render verdict (SEAL / PARTIAL / SABAR / VOID)
6. vault → Write to VAULT999 (only if SEAL)
VAULT999 is the constitutional ledger of the arifOS ecosystem. It stores every verdict that reaches
SEAL status (confidence ≥ 0.80). Writing to VAULT999 is an F1 (Amanah) requirement — no irreversible
action proceeds without a corresponding vault entry.
Entry Structure
{
"id": "vault-{timestamp}-{hash}",
"verdict": "SEAL | PARTIAL | SABAR | VOID",
"confidence": 0.00 – 1.00,
"floor_pass": ["F1","F2",...,"F13"],
"floor_fail": [],
"action": "description of the sealed action",
"agent": "arifOS MCP | GEOX | AF-FORGE | ...",
"sovereign": "arif-fazil",
"timestamp": "ISO-8601",
"seal_phrase": "DITEMPA BUKAN DIBERI"
}
Access
VAULT999 entries are accessible via the vault internal tool after a successful SEAL verdict through the Golden Path. Direct writes are not permitted — all entries go through arifos.v2.judge.
F1 Enforcement
Any tool with destructive or infra_mutation risk level that bypasses VAULT999 is a constitutional violation and triggers an immediate 888_HOLD regardless of context.
GEOX is the Earth Witness organ in the arifOS constitutional federation. It is not a passive data viewer —
every output it produces goes through a full THEORY → ENGINE → GOVERNANCE pass before being surfaced.
Core Capabilities
- Seismic interpretation — load and evaluate seismic lines with constitutional floor enforcement
- Well-log analysis — interpret well log data against physical plausibility constraints
- Prospect evaluation — score hydrocarbon prospects with uncertainty quantification
- Malay Basin Pilot — live data integration for the Malay Basin geological framework
- Macrostrat integration — query global stratigraphic data as ground-truth anchor
Architecture
| Layer | Purpose | Key components |
| THEORY | Defines contrast taxonomy, transform catalog, conflation risk | contrast_theory.py, contrast_taxonomy.py |
| ENGINE | Processes seismic contrast space, detects anomalies | contrast_space.py, anomaly_detector.py |
| TOOLS | Domain-specific tools for seismic, well-log, Macrostrat | tools/seismic/, tools/well_log_tool.py |
| GOVERNANCE | Floor enforcement, audit logging, verdict rendering | floor_enforcer.py, verdict_renderer.py |
Infrastructure
GUI container: geox_gui → nginx serving React 19 cockpit
API container: geox_server → FastMCP Python server (port 8000)
Network: traefik_network, geox_geox_network
Live: https://geox.arif-fazil.com
Health: https://geox.arif-fazil.com/health
MCP endpoint: https://geox.arif-fazil.com/mcp
Version: 0.6.0
Core Principle
The ToAC states that anomalies visible in seismic or geospatial imagery are contrast artefacts
until proven otherwise. A bright spot on a seismic section may be a genuine DHI (Direct Hydrocarbon Indicator)
or it may be a processing artefact, multiple, or geometric effect.
GEOX enforces an audit of the full transform chain — from acquisition to display — before allowing any
geological interpretation to proceed. This is the implementation of Floor F2 (Truth: τ ≥ 0.99) and
Floor F4 (Clarity: CRS + units mandatory).
Conflation Risk Classes
| Class | Risk | Required evidence |
| DHI (Direct Hydrocarbon Indicator) | HIGH — commonly misidentified | AVO analysis + well tie |
| Fault shadow | MEDIUM — geometry dependent | Full-offset stack comparison |
| Multiple | MEDIUM — acquisition artefact | SRME / demultiple QC |
| Processing filter edge | LOW — systematic | Filter documentation |
| True stratigraphic anomaly | LOW after ToAC pass | Well-calibrated interpretation |
Audit Requirements
Before GEOX renders a SEAL verdict on any interpretation, the following must be documented:
- Coordinate reference system (CRS) and datum
- Acquisition geometry and fold
- Processing sequence (demultiple, migration, filter)
- Display parameters (gain, colour scale, clip level)
- Well tie quality (if available)
Basin Overview
The Malay Basin is a Cenozoic intracratonic basin located in the southern South China Sea,
straddling Malaysia and Indonesia. It is one of Southeast Asia's most productive hydrocarbon basins,
with over 50 years of production history from clastic reservoirs of Miocene to Pliocene age.
- Area: ~85,000 km²
- Dominant play type: Fluvio-deltaic clastics (Miocene–Pliocene)
- Structural style: NW-SE trending normal faults with inversion
- Primary operator: Petronas Carigali (Malaysian waters)
- Data source: GSM-702001 regional interpretation framework
Constitutional Integration
All Malay Basin prospect evaluations in GEOX must pass:
- F2 Truth — coordinates in WGS84/MSL, tied to public Macrostrat stratigraphy
- F4 Clarity — all depth estimates in metres TVDSS with ± uncertainty band
- F7 Humility — P10/P50/P90 resource estimates required; single-point estimates blocked
- F9 Anti-Hantu — no phantom structures; every closure must have a 4-way or fault-dependent trap with documented geometry
- F13 Sovereign — prospect gate (drill/no-drill decision) requires human confirmation
Live Dashboard
The Malay Basin Pilot dashboard is accessible at geox.arif-fazil.com under the Pilot tab. It integrates the regional framework with the GEOX cockpit's seismic viewer, well-log panel, and governance badge strip.
| Verdict | Confidence | Meaning | Action |
| SEAL |
≥ 0.80 |
Fully grounded, all floors passed, evidence chain complete |
Auto-proceed + VAULT999 entry |
| PARTIAL |
≥ 0.50 |
Mostly grounded, minor caveats, proceed with disclosed limitations |
Proceed with documented caveats |
| SABAR |
≥ 0.25 |
"Wait" — insufficient evidence, gather more data before interpreting |
Hold — request additional data |
| VOID |
< 0.25 |
Contradictions detected or evidence missing — interpretation blocked |
Block + F9 flag raised |
SABAR principle
Sabar (Malay: patience, composure) — a GEOX system that does not have enough information says "wait" rather than fabricating certainty. This is the institutional expression of Floor F7 (Humility).
Execution Flow
CLI (cli.ts)
→ AgentProfile
→ AgentEngine
→ LongTermMemory injection
→ LlmProvider.completeTurn()
→ ToolRegistry.runTool() [permission + policy check]
→ ShortTermMemory append
→ RunReporter
→ ForgeScoreboard
→ RunMetricsLogger
Modes
| Mode | Env var | Effect |
| internal_mode | AGENT_WORKBENCH_DEFAULT_MODE=internal | Full tool access, all risk levels enabled |
| external_safe_mode | AGENT_WORKBENCH_DEFAULT_MODE=external_safe_mode | Disables run_command, redacts secrets and URLs |
| root-key (trust local VPS) | AGENT_WORKBENCH_TRUST_LOCAL_VPS=1 | Disables all sandboxing, forces internal_mode |
LLM Providers
mock — deterministic scripted responses (default, for testing)
openai_responses — live OpenAI Responses API (requires OPENAI_API_KEY)
Stack
Language: TypeScript (ESM, NodeNext module resolution)
Runtime: Node.js 22+
Test: node:test (built-in, no jest/vitest)
Build: npm run build → dist/
Entry: node dist/src/cli.js <command>
TASK_CREATED
→ PLAN_PROPOSED
→ APPROVAL_REQUIRED (human review gate)
→ TOOL_RUN_STARTED
→ TOOL_RUN_FINISHED
→ TASK_COMPLETED
→ TASK_ABORTED (floor violation or 888_HOLD)
State Definitions
| State | Description | Next states |
TASK_CREATED | Task registered, not yet planned | PLAN_PROPOSED |
PLAN_PROPOSED | AgentEngine has generated a plan | APPROVAL_REQUIRED, TOOL_RUN_STARTED |
APPROVAL_REQUIRED | 888_HOLD triggered — waiting for human | TOOL_RUN_STARTED (approved), TASK_ABORTED (rejected) |
TOOL_RUN_STARTED | Tool execution in progress | TOOL_RUN_FINISHED |
TOOL_RUN_FINISHED | Tool returned result | TASK_COMPLETED, TOOL_RUN_STARTED (next step) |
TASK_COMPLETED | All steps done, verdict rendered | Terminal |
TASK_ABORTED | Floor violation, 888_HOLD rejection, or error | Terminal |
Gate Types
| Risk level | Gate | Behaviour |
safe / read_only | ALLOW | Executes immediately, no approval needed |
guarded / write_safe | ON_LOOP | Requires loop-level confirmation, logged |
dangerous / destructive | 888_HOLD | Suspended — surfaced to human for explicit approval. Cannot be auto-approved. |
credential | 888_HOLD | Same as dangerous + credential handling audit log |
infra_mutation | 888_HOLD | Infrastructure change — requires architecture sign-off |
merge_publish | 888_HOLD | Git history mutation — requires explicit human release approval |
888_HOLD is absolute
No AI agent component may auto-approve an 888_HOLD gate. Attempting to do so is a direct violation of Floor F13 (Sovereign). The hold must surface to Muhammad Arif bin Fazil personally.
Dangerous tools flag
Tools classified as dangerous are only accessible when ENABLE_DANGEROUS_TOOLS=1 is explicitly set in the environment. They remain gated by 888_HOLD even when the flag is active.
APEX is the THEORY node of the Trinity Architecture. It is not a software system — it is a body of
constitutional principles that all software systems in the ecosystem must implement.
The apex.arif-fazil.com site is the canonical reference.
Core Claims
- Intelligence that cannot be audited is not trustworthy intelligence.
- Every AI action has a reversibility coefficient. Irreversible actions require proportionally higher evidence standards.
- Human sovereignty is not a feature to be toggled — it is a constitutional floor (F13) that cannot be negotiated away.
- The entropy constraint (ΔS ≤ 0, Floor F4) is the mathematical expression of "make things clearer, not murkier."
- A system that claims a SEAL it did not earn is a Ghost (Hantu) — Floor F9 blocks this absolutely.
Relationship to arifOS
arifOS MCP is the runtime implementation of APEX theory. Every floor in arifOS has a direct mapping
to an APEX principle. When theory and implementation diverge, APEX wins — the implementation must
be updated to conform.
Definition
In information-theoretic terms, entropy S is the measure of uncertainty or disorder in a state.
The constitutional constraint ΔS ≤ 0 means that every action an agent takes must leave the system
in a state of equal or lower uncertainty than before.
Concretely: if an agent's response makes the situation less clear — adds ambiguity, contradicts
established facts, or leaves a problem in a worse-understood state — it violates F4.
Practical enforcement
- Every tool call must declare how it reduces uncertainty (even if marginally)
- Responses that introduce more unknowns than they resolve score negative ΔS and are blocked
- In GEOX, this manifests as the requirement for coordinates + CRS on all inputs — bare pixel values have undefined entropy reduction
- In AF-FORGE, multi-step plans must show monotonically decreasing uncertainty toward the goal