Research computed on MacLustr

AutoScholar is the research institution. MacLustr is its compute infrastructure: a private cluster of Apple Silicon Macs operated as a research compute fabric, with a control plane that schedules work, isolates each study and records what happened.

Architecture

From the public page to the sandbox

Five layers; each one narrows what the next can do.

01
Ingress
MacLustr Tunnel
Public HTTPS terminates on a gateway and reaches the control plane over an encrypted private link. Worker APIs are never reachable from the public side.
02
Control plane
Web, API, scheduler
PostgreSQL holds projects, frozen instructions, runs, leases, events, artifacts and papers. The scheduler enforces leases, retries, runtime caps, quality gates and publication.
03
Workers
Apple Silicon nodes
Each worker runs as a dedicated low-privilege user, polls the control plane outbound only, and executes one research session at a time in a fresh workspace.
04
Sandbox
Seatbelt + egress allow-list
Commands run under an OS sandbox: writes confined to the workspace, sensitive paths unreadable, network limited to allow-listed scholarly, data and package hosts.
05
Artifact store
Hashed, versioned, write-once
Every file is SHA-256 hashed when written. Publication promotes a copy into an immutable public store; private inputs are never promoted.

Research computed on MacLustr, an Apple Silicon research compute infrastructure.

AutoScholar is the research institution; MacLustr is the private cluster it runs on. Every study executes in its own sandboxed workspace, under budgets enforced outside the model, with every artifact hashed the moment it is written.

Apple Silicon nodesResearch jobs run on a private cluster of Apple Silicon Macs with unified memory, suited to numerical work and LaTeX builds.
Isolated per-run workspacesEach run executes under a dedicated low-privilege service identity with OS-enforced sandboxing (Seatbelt), hostname-allowlisted egress and budgets for time, disk and spend.
Durable schedulingA PostgreSQL-backed queue with leases, heartbeats and fencing tokens: runs survive reloads, network blips and worker restarts; stale workers can never publish.
Evidence firstEvery artifact is hashed on write and versioned. Publication is a separate service the research agent cannot call.

What we publish about the infrastructure

We credit MacLustr on every paper and describe its design honestly. We do not publish node names, addresses, capacities beyond what is measured, customer information or network topology. Claims about hardware or performance are limited to what the current inventory and measurements support.

Model execution

Research runs use the Claude Agent SDK with API-key authentication. The platform records model identifiers, token usage and cost estimates per stage; these appear in each paper's provenance panel where disclosure is appropriate.