Kaira Platform

Kaira is the long-running system behind the rest of this work: an autonomous engineering platform that takes a unit of work, runs it on real hardware, and refuses to call it done until the required check has actually run, passed and been persisted. Evidence, not assertion — that rule is what most of the design exists to enforce.

It is built architecture-first. Three architecture documents and a red-team review are the standing authority, held in version control with their amendments recoverable as commits, and they outrank any code that drifts from them. Underneath sits a strict authority order: what is measured to be true beats what policy allows, which beats what the architecture says, which beats what any model proposes.

The runtime separates concerns that are usually one blob. A scheduler and task state machine own execution; typed capabilities and a registry own what can be invoked; providers put local and hosted models behind a single contract; a resource governor hands out leases against cgroups and GPUs; execution happens in git worktrees inside a sandbox. Guardian — deterministic, model-free — owns lifecycle, health, the release switch and rollback, and runs below and outside the runtime it protects, so a bad release cannot talk its way out of being reverted.

Secrets are redacted at the capture boundary before anything is persisted, ordinary workers can never write a stable branch, and events leave through a transactional outbox rather than a best-effort publish. The platform runs on a dedicated machine on a private tailnet and is being taken through its self-hosting path — the point at which it maintains its own codebase.