A true agentic OS. A new architecture for agent execution, built on operating-system principles, with the harness as a path toward superintelligence.
Give a local model one bounded job. Keep the final decision in Copilot.
Sev Arabic Preview v0.1 — Arabic-first typed decision engine (research preview). Runtime code; model on Hugging Face (gated).
Evidence-bound Harness-of-Harness for verifiable long-horizon agent engineering: receipt-bound acceptance, independent QA, durable program state, recovery, explicit limitations. Pre-release v0.3.0a2 (stable: v0.1.0), developed under its own orchestration since 2026-09-07, with product-run merges since 2026-09-25. Research preview.
Native C17 language, deterministic QBC/QVM runtime, virtual QCPU, and approval-first security research preview.
Research preview of Closure Engineering: testing whether coding agents need a project-completeness layer over task graphs.
Research preview of execution contracts for portable placement across client, edge, and server.
A gravity-inspired physics optimization kernel with a constraint-driven runtime (Ising/QUBO as the description layer).
System-One decision intelligence model: answers all typed questions in a single inference pass with calibrated probabilities. Built from scratch in PyTorch.
Research preview: timing-aware turn-taking for spoken language models with external Time Adapters and action-token control.
Publication-safe research preview of a Vietnamese-first small code model with execution verification and Blackwell experiments.
Experimental local-first evidence retrieval backend for AI agents and local LLM workflows.
Alpha / research preview: local-first verifier for coding-agent patches (modeled runtime-failure analysis, behavioral contracts, differential replay, MCP) for Python
To associate your repository with the research-preview topic, visit your repo's landing page and select "manage topics."