MD-OS

AGI Paradigm

MD-OS (Artificial Prefrontal Cortex) v5.0 should not be described as AGI.

It should be described as a Markdown-native Operating Filesystem for AGI-like agents and robotic systems.

More directly, it is the natural-language agentic layer between MD-OS and the real substrates that a host machine exposes: OS, hardware, applications, services, devices, robots, sensors, and actuators.

The important practical paradigm is natural-language robotic-agentic programming: using durable language artifacts to program the operating behavior of a complex ecosystem of agents, tools, work systems, devices, sensors, robots, policies, telemetry, approvals, and recovery paths.

Core Position

MD-OS is not AGI.
MD-OS is the Markdown-native Operating Filesystem that persistent,
inspectable, tool-using intelligence needs to act across sessions and recover
from failure.

The system does not provide intelligence by itself. A model or host runtime provides reasoning. MD-OS provides the durable operating surface around that reasoning:

• persistent memory
• readable goals
• project and task state
• bounded connector actions
• telemetry snapshots
• append-only audit history
• deterministic rebuilds
• replayable recovery
• human correction through normal files

The executable learning direction is a model-agnostic cognitive hypervisor:

models = replaceable processors
connectors = bounded drivers
Markdown = human and constitutional source
Cognitive IR = executable transaction contract
episodes = real traces
skills = evaluated reusable programs
verifiers = independent postcondition checks

The canonical first path is the Cognitive Transaction Loop in md-os/kb/COGNITIVE_TRANSACTION_LOOP_MODEL.md. The historical cortex agi command remains a compatibility alias and must not be described as a separate AGI layer.

The first measurable vertical is the software-repair benchmark defined in md-os/kb/SOFTWARE_REPAIR_BENCHMARK_MODEL.md. It now compiles typed PlanGraphs through an append-only CandidateProvider boundary before verification. The controlled provider validates planning, provenance, diversity, and isolation only. A separate bounded learning provider now supplies the first narrow cross-instance measurement: after two independently verified development episodes, one induced parameterized skill is evaluated on a distinct validation case and two sealed holdouts at the same one-attempt budget used by the no-skill baseline. The reference run improves holdout success from 0/2 to 2/2 with no detected contamination or regression. This supports learning and transfer only inside one declared four-constraint hypothesis family. A separate bounded solver-transport experiment and cognitive-unity fixture now test explicit frame transformations in authored finite families. Broad open-domain invention, general cross-domain competence, continual autonomous improvement, and AGI remain unmeasured.

The executable and epistemic contract is defined in md-os/kb/NEUROMORPHIC_LEARNING_ACCELERATOR_MODEL.md.

Cross-Domain Cognitive Unity Research Path

The candidate principle is not that one solver, model, or tensor is AGI. Generality is studied as the verified ability to construct transformations between declared frames while preserving the relations required by the task. Cognitive unity is the persistent causal integration of those frames, transformations, invariants, goals, memories, actions, and evidence into one governed decision process.

APFC is the controller and functional intelligence extender in this model. It makes the recurrence observable across bounded host-model calls, generates competing candidate laws, protects development/target separation, schedules falsification tests, and reuses only evidence-bound results. When an external representation has declared axes and bases, finite tensor operators can express the candidate transformation. Cortex does not inspect or modify neural hidden layers; it extends their operational reach through persistent external state.

The implemented fixture verifies one finite law family and fail-closed promotion path. It does not prove that this principle is sufficient for AGI, that the mind is one tensor, or that governance telemetry is consciousness. The canonical formulation and falsifiers are in md-os/kb/CROSS_DOMAIN_COGNITIVE_UNITY_MODEL.md.

Mental Model

Model
  reasoning engine

Prompt
  momentary instruction

Context window
  short-term attention

Tools and connectors
  hands and sensors

MD-OS
  persistent Operating Filesystem, action ledger, agenda, telemetry, audit, replay

Robot Stack Position

MD-OS should sit above the safety-critical control loop.

Human intent
  -> MD-OS
  -> robot connector
  -> ROS 2 / robotics stack / device API
  -> bounded action
  -> telemetry snapshot
  -> Markdown/JSON state
  -> replayable audit trail

MD-OS should not replace:

• firmware
• motor control
• emergency stops
• real-time motion planning
• hardware drivers
• safety-certified systems

MD-OS can coordinate:

• mission goals
• plans
• task state
• robot observations
• action proposals
• approvals
• post-action audit
• recovery paths

Raspberry Pi Example

A Raspberry Pi is a concrete physical host for this pattern:

Raspberry Pi hardware
  -> Raspberry Pi OS / Linux
  -> GPIO, serial, camera, MQTT, HTTP, shell, or ROS 2 interface
  -> bounded MD-OS connector
  -> snapshot, project state, agenda, journal, and continuity files

MD-OS on a Raspberry Pi can supervise local automation, sensors, monitoring, or lightweight robot workflows while Linux and device-specific libraries continue to handle hardware access and safety-critical execution.

Safe Claim Language

Use:

MD-OS is a Markdown-native Operating Filesystem for persistent AGI-like agents
and robotic systems.

Use:

MD-OS gives AGI-like agents an inspectable filesystem for memory, goals,
actions, telemetry, audit, and recovery.

Avoid:

MD-OS is AGI.
MD-OS creates AGI.
MD-OS replaces ROS, Linux, firmware, or safety systems.

This site is open source. Improve this page.