• v0.13–14Persistence — Blake3 write-ahead log, crash recovery: apply → reboot → state intact
• v0.15–16Trust & contracts — three-stage Ed25519 signature chain; two keys become law
• v0.17AI workloads — in-cell inference, chat API, MCP tool cells
• v0.18Preemption & multicore — time slices, four cores, exclusive core ownership as a capability
• v0.19Observability & audit — per-cell CPU time, log rings, request tracing, the tamper-evident audit chain
• v0.20SDK & app lifecycle — signed app bundles, one-command install, the OS as a signed CI-built image
• v0.21Data protection — encryption at rest, a net-free keystore, credentials an app never holds
• v0.22ACP Console — browser login with roles, RBAC + CSRF on every write, lifecycle, topology, deploy with diff, value-free secrets
• v0.23HA replica groups — replicas: N, rolling restarts, gate: never 0 replicas alive
• v0.23Service routing — round-robin in the gateways; 3 rollouts under fire, 802 req, 0 errors
• v0.23Boot latency — unattended convergence in 7.4 s, per-phase boot trace
• v0.23Benchmark vs. Linux (TCG) — measured, release builds, QEMU/TCG on both sides: p50 0.59× Linux on the no-IPC path, 0.68–1.44× with IPC; first answer 3.4 s vs. 6.6 s. Emulated on both sides — and the syscall and IPC ratios from this run turned out to be emulation artefacts, corrected by the KVM run below
• newFirst boot on KVM — Arkane runs on real virtualization (AMD host). Three kernel bugs that TCG had waved through were found and fixed on the way: IDT gaps, MSI-X swapgs, SS.RPL selector
• newFirst measurements on real hardware-assisted virt — Arkane and Linux timed on the same host, same run: syscall p50 224 cycles vs. 1393 (getpid, ~6× faster), IPC round-trip 36850 vs. 40864 (AF_UNIX, ~10% faster). One host, dev profile, n=2 — a first datapoint, not a benchmark suite

• doneIPC cost decomposition — measured before optimising: copies and tick-waits ruled out (4.4%), the expensive block is park + return through the run loop; two hypothesised cheap fixes were built, measured at −0.6% / −0.0%, and reverted rather than shipped as improvements. The KVM run then reframed the whole question: the path it was meant to rescue is already faster than AF_UNIX
• v0.24Benchmark & SDK developer experience — the measurement work above lives here, alongside SDK ergonomics: arkctl new builds out of tree, arkctl manifest fill writes the image hash for you, and the common failure paths explain themselves; remaining ergonomics under review. The serving-path decomposition landed here too: 80–89% of the first KVM serving latency was our own per-request debug serial output, and with it off Arkane serves at 0.98–1.34× Linux p50, with parity on the public tree
• v0.25ArkDB — planned, and the plan is committed: a database built with the SDK, as a contracted cell, with golden tests against SQLite and every performance claim gated on a benchmark. Nothing about it is claimed yet — this line describes a plan, not a product
• laterMulti-node — replica groups spread across Arkane nodes; today HA means one machine
• post-1.0Bare metal & hardware trust — IOMMU, TPM, real iron; boot under 3 s. Version 1.0 means the finished AI OS proven on QEMU; hardware comes after. The KVM boot is the bridge pier on the way there, not the arrival

The target for 1.0 is still QEMU — bare metal is a named trigger, not a promise. KVM does not change that line; it moves the system one pier closer to it, on hardware that executes the kernel for real. That sentence is on this page on purpose.