Publication Model: Connector OS is released as a reference architecture. Issues may be opened for technical clarification. Open-ended community discussion is intentionally not enabled.
Version: Architecture v1.0 (Whitepaper Published February 2026)
Tagline: Intelligence is not only in the model - it is in the architecture.
A modular human–AI systems architecture grounded in adaptive control, threshold regulation, and closed-loop stability.
This repository is summarized in the article:
Architecture Before Scale: A Control-Theoretic Framework for Stable AI Systems
Published February 20, 2026
→ Read the article on LinkedIn
The article provides a narrative overview of the architectural principles. This repository contains the full technical specifications and implementation details.
Connector OS is a layered architecture for building state-aware AI systems using established control-theoretic principles.
Modern AI deployments optimize prediction.
Connector OS adds the missing layer: explicit regulation.
The central thesis:
Model capability is necessary but insufficient.
System stability is an architectural property.
Rather than treating AI models as standalone agents, Connector OS treats them as pluggable components inside a regulated control stack.
The focus is coordination, routing, thresholding, and feedback - not scaling model size.
Connector OS is published as a reference architecture.
It is intended to be studied, adapted, stress-tested, or reinterpreted within other system designs.
Many components are modular by design and can be extracted independently of the full stack.
If you are evaluating, cloning, or building variations locally - that is its intended use case.
Attribution is appreciated but not required.
The goal is structural legibility and reuse.
This repository provides:
- An 8-layer architectural stack for regulated human–AI interaction
- Layer 2 (CMP): Context compression under bandwidth constraints
- Layer 3: Explicit control logic (thresholds, hysteresis, routing)
- Layer 5: Human state modeling as a system variable
- Minimum Viable Modules (MVMs) demonstrating practical implementations
- Stress-test experiments validating architectural behavior under constraint
- Cross-domain validation from hydrology, power systems, and physiology
This is not an AGI proposal.
It is a control-systems framework for stabilizing AI deployments.
Connector OS is structured as an 8-layer stack:
Layer 7: Co-Thought (Human+AI joint reasoning)
Layer 6: AI Models (pluggable brains)
Layer 5: Human State Loop (bio/affective bands)
Layer 4: Actuators (lights, sound, haptics, UI)
Layer 3: Control Logic (dams, grids, feedback)
Layer 2: Context Map Protocol (CMP glyphs)
Layer 1: Sensors (HRV, gaze, voice, input devices)
Layer 0: F₀ Resonance (shared timing / 40 Hz band)
Shared timing base (~40 Hz reference band) for aligning biological and digital update cycles in closed loops.
Raw input streams: HRV, interaction cadence, voice prosody, environmental data.
Transforms raw streams into structured state glyphs.
Preserves semantic structure under bandwidth constraints.
Implements threshold-based regulation:
- Spillway logic (controlled release)
- Load balancing
- Hysteresis
- Adaptive routing
This layer governs when and how strongly models are invoked.
All output channels: UI, voice, haptics, environmental modulation.
Models the human as a dynamic system:
- Stress
- Cognitive bandwidth
- Abstraction tolerance
- Fatigue
Frontier or local models treated as interchangeable plugins.
Closed-loop interaction state where adaptive regulation supports joint reasoning.
Detailed specification:
docs/02_layered_architecture.md
Connector OS encodes three universal regulatory mechanisms:
Systems must sense their own output and adjust accordingly.
Intervention occurs only when defined boundaries are crossed.
System behavior scales with human state and infrastructure constraints.
These principles appear across:
- Dams (spillways)
- Power grids (load redistribution)
- Physiology (homeostasis)
Connector OS applies them to AI system coordination.
See: docs/04_control_laws_and_analogies.md
Connector OS is modular. It is not deployed as a monolithic system.
A buildable module that:
- Reads HRV
- Computes state deviation
- Applies threshold logic
- Modulates environment (lights, haptics)
Spec:
mvm/MVM-1_vibe-check_prometheus-1.md
- MVM-2 — Shadow State Assistant
- MVM-3 — Haptic Ticker
- MVM-4+ — Experimental state-aware modules
If you want to see Connector OS principles in practice immediately:
src/shortcut_recipes/prometheus-1_apple-shortcuts.md
This implementation demonstrates:
- Layer 1 — Sensor input (HRV via HealthKit)
- Layer 2 — State normalization (baseline deviation)
- Layer 3 — Threshold logic (15% / 30% bands)
- Layer 4 — Actuation (lights, haptics, UI)
- Layer 5 — Closed-loop human state regulation
It does not require a language model.
It demonstrates deterministic control logic under real physiological input.
This is the minimal reproducible example of the architecture.
The formal architectural treatment is available here:
Architecture Before Scale: A Control-Theoretic Framework for Stable AI Systems
📄 PDF (Citable Artifact):
whitepaper/Architecture_Before_Scale_v1.0.pdf
📝 Markdown Source:
whitepaper/Architecture_Before_Scale_v1.0.md
📘 Versioning Policy:
whitepaper/whitepaper_VERSIONING.md
This document presents:
- Formal problem statement
- Universal control law grounding
- Layer-by-layer architectural specification
- Stress-test validation
- Governance implications
Whitepaper status: Final (v1.0, February 2026)
Connector OS is evaluated under infrastructure stress:
Simulates:
- Narrow output pipes
- Latency spikes
- Queue overload
Demonstrates that architectural regulation preserves coherence where naive stacks degrade.
See:
experiments/EXP-01_bandwidth_constraint_test.md
Connector OS is designed as a bounded regulatory layer.
- Explicit thresholds
- Inspectable logic
- Nudge-first actuation
- User override always available
- Relative baselines (no global absolutes)
This is regulation architecture, not behavioral manipulation.
See: docs/06_ethics_and_guardrails.md
This repository is for:
- Systems engineers
- Control theorists
- AI infrastructure architects
- Researchers in adaptive regulation
- Builders of multimodal or embodied AI systems
connector-os/
├── README.md
├── docs/
│ ├── 01_overview_connector_os.md
│ ├── 02_layered_architecture.md
│ ├── 03_signal_topography.md
│ ├── 04_control_laws_and_analogies.md
│ ├── 08_cross_domain_validation.md
│ ├── glossary.md
│
├── mvm/
│ └── MVM-1_vibe-check_prometheus-1.md
│
├── experiments/
│ └── EXP-01_bandwidth_constraint_test.md
|
├── whitepaper/
│ ├── Architecture_Before_Scale_v1.0.pdf
│ ├── Architecture_Before_Scale_v1.0.md
│ ├── whitepaper_README.md
│ ├── whitepaper_VERSIONING.md
│ └── figures/
└── meta/
└── contributor_models.md
- Architecture: v1.0 (Whitepaper Published)
- MVM-1: Implementable
- Additional MVMs: Iterative
- Whitepaper: Archived release artifact (v1.0)
MIT License — open, forkable, extensible.
Primary Architect: Zee / Leena Thomas
System Design & Coherence: Thea
Model-assisted documentation and diagrams credited in meta/contributor_models.md
Stability is not a property of intelligence alone.
It is a property of regulated systems. "The intelligence is in the connectors."
This repository addresses the "Body Problem" for AI - how to give stateless models state-awareness and stability.
For a complete catalog of related research:
📂 AI Safety & Systems Architecture Research Index
Thematically related:
- Voice Mode Forensics - Prosodic alignment failures that informed this architecture
- Embodied Agent Governance - Governance patterns for agents with bodies
- The Continuity Problem - Why state persistence requires governance
- Designing for Failure - Pattern language for catastrophic-state recovery discipline and survivability architecture.