SpacePilot · Design Paper · August 2026

SpacePilot: A Provenance-First Decision Layer for Single-Owner AI Compute Fleets

MotionVector · draft v0 · PDF, 7 pages

Status — design paper. A single-machine core is implemented and running; the multi-machine design is presented as design, not as evaluated results. Numbers marked ⟨TBD⟩ are placeholders pending measurement.

Read the PDFDownload

Abstract

Model and runtime optimizations are well served by open source; the layer that decides where a workload should run is not. Production schedulers assume datacenter physics — nodes that are up or failed, specifications that predict performance, durable cluster state. None of the eleven systems we survey models a suspended consumer machine as a schedulable state. Yet the fastest-growing deployment target for open models is the single-owner fleet: a laptop, a workstation, occasionally rented cloud machines, and managed inference APIs, all owned or paid for by one person or small team.

We present the design of SpacePilot, a decision layer for such fleets, built on three commitments. Provenance is typed: every performance number is labeled flown (measured on this fleet), on paper (cited and dated, external), or unflown (no evidence), and the type system refuses to blend them. Readiness beats specification: placement weighs which machine already holds the weights in memory, not rated FLOPS, via a cost model that prices cold starts explicitly. Sleep is a first-class state: a suspended machine is asleep, not gone; work may wait for it, and the interface says so. Fleet state divides into three classes by reconciliation semantics — an append-only measurement log that merges by union, a single-author signed order set, and an observed live picture that is never persisted — which removes the need for consensus in a fleet with one owner.

In the paper

© 2026 MotionVector · spacepilot.dev · You decide what to run. SpacePilot decides how and where.