Cohesix 1.2.0 Ships Governed Edge AI Workflows

Summary

Cohesix 1.2.0 is a release for running edge AI jobs through one governed workflow: a client submits a job, the Queen running in seL4 on QEMU or Raspberry Pi 4 admits it, a selected host executes it, and the resulting provider output can be inspected. Admission acknowledgement is kept separate from CUDA or MLX completion. If a connection is lost, Cohesix retains the original job identity and reconciles the outcome instead of silently replaying the work; failed or uncertain candidates remain visible. GPU libraries stay on the host that owns the accelerator: the supported NVIDIA Jetson path uses CUDA for inference and adapter work, while Apple Silicon uses MLX with Apple’s Metal runtime. Provider selection is explicit, so jobs do not quietly move between MLX, CUDA, or CPU. SwarmUI, the Cohesix CLI and Python clients, MCP tools, and A2A agents can use the same governed job, with MCP and A2A enabled by default in the QEMU and Pi gateway profiles. The PEFT workflow connects training or import with evaluation, canary deployment, promotion, and rollback, while rejected candidates remain rejected even if an earlier serving version is restored. Release artifacts cover Apple Silicon, JetPack Linux ARM64, Raspberry Pi 4, QEMU, and Python clients, including `cohesix==1.2.0` on PyPI. The documented references are JetPack 7.2.1/L4T 39.2.1 on ARM64, Apple Silicon, and wired networking on Raspberry Pi 4; other platform and Wi-Fi combinations require separate qualification. Installers do not download models or copy credentials, and the release advises verifying checksums and signatures before installation.