🧠 coreason-runtime

The official zero-trust, high-throughput kinetic execution engine for the coreason-manifest ontology.

coreason-runtime is a State-of-the-Art (SOTA) 2026 cybernetic execution engine. It abandons legacy, fragile "chain-of-thought" LLM scripting in favor of deterministic Active Inference, Topological Data Analysis (TDA), and strictly bounded Markov Decision Processes. It is the definitive implementation of the CoReason Tripartite Doctrine for Tier-1 Kinetic Execution.

If coreason-manifest is the DNA of your multi-agent topologies, coreason-runtime is the biological cell that safely executes them.

---

🚀 The Paradigm Shift

Modern enterprise AI cannot rely on unbounded while True loops and raw Python exec(). The coreason-runtime enforces mathematical rigor at every boundary:

• Deterministic Orchestration: Built on Temporal, Cognitive Topology executions are durably serialized. If a GPU dies or a network request fails, the topology pauses, rehydrates, and resumes exactly where it left off. No amnesia. No ghost processes.
• Zero-Trust WASM Sandboxing: Kinetic actions (Tools) are executed inside isolated WebAssembly environments via Extism. Agents can execute complex IO without ever touching your host's root kernel or filesystem.
• Epistemic Vector Ledger: Native, zero-copy integration with LanceDB. The runtime automatically projects the agent's latent state into an embedded vector memory layer.
• Embedded Medallion Analytics: No need for heavy Spark clusters. Raw telemetry (SSE) is ingested via dlt and transformed into Silver/Gold analytical intelligence matrices using Rust-backed Polars directly inside the daemon.
• Human-in-the-Loop (HITL) Webhooks: When an agent calculates high Variational Free Energy (epistemic uncertainty), it durably suspends its thread and emits an Oracle Request, waiting safely for a human expert to inject resolving priors via API.
• Bipartite Proposer-Verifier Protocol: The runtime is physically isolated from local OS capability generation. To fabricate assets, the runtime strictly proposes topological models over air-gapped MCP boundaries to the remote Universal Asset Forge (coreason-meta-engineering).

---

⚡ Installation

We utilize uv for ultra-fast, deterministic resolution. Ensure you are running Python 3.14+.

uv add coreason-runtime

Note: For bare-metal enterprise deployment with SGLang GPU passthrough, refer to our Docker Deployment Guide.

---

🐳 Docker Deployment & Hardware Capabilities

The local docker development environment dynamically scales its container footprint depending on the capabilities of the host system.

If the host system does not have an NVIDIA GPU, it skips the heavy sglang (Cognitive Engine) image download (10GB+) and installs a lightweight runtime package without CUDA/PyTorch dependencies.

To automatically configure your environment based on host capabilities, run:

Windows (PowerShell)

./scripts/detect_capabilities.ps1

macOS/Linux

chmod +x ./scripts/detect_capabilities.sh
./scripts/detect_capabilities.sh

These scripts will automatically inspect your hardware and generate/update the local .env configuration file with:

• COMPOSE_PROFILES=gpu (activated only if NVIDIA GPU is detected)
• EXTRAS=inference (packages deep learning libraries only if GPU is present)

After running the capability check, start the local stack:

docker compose up --build

---

🛠️ Quickstart

The runtime is designed to be operated via its CLI or mounted as an API edge.

1. Run a Local Cognitive Topology

To execute a mathematically verified agentic topology, simply pass the JSON/YAML manifest to the runtime:

coreason run ./my_topology_manifest.json

2. Boot the API Edge & Telemetry Broker

To boot the runtime as a continuous daemon (exposing the CRDT State Sync, Schema Projection, and Server-Sent Events telemetry):

coreason serve --port 8000

Your frontend IDE can now connect to http://localhost:8000/api/v1/telemetry/stream to visualize the active inference loops in real-time.

---

🏗️ Architecture

The runtime operates across five isolated computational boundaries under the CoReason Tripartite Doctrine:

1. The Orchestrator: Temporal Python SDK running deterministic AST-scanned workflows.
2. The Cognitive Engine: SGLang routing for sub-millisecond constrained tensor inference.
3. The Kinetic Sandbox: Extism executing .wasm tools with zero-trust lattices.
4. The Epistemic Store: LanceDB & Polars managing long-term vectors and ETL metrics.
5. The Universal Asset Forge: A decoupled MCP channel connecting strictly to the coreason-meta-engineering Fabrication Lines to physically synthesize assets via the Bipartite generation pipeline.

For a deep dive into the cybernetic loop, read the Architecture Documentation.

---

📜 License

This software is proprietary and dual-licensed under the Prosperity Public License 3.0. Commercial use beyond a 30-day trial requires a separate commercial license. See the LICENSE file for details.

Copyright (c) 2026 CoReason, Inc.