Useful AI still means renting reasoning, memory, and tools from a few companies most users cannot afford to leave. Capable smaller models now run on a laptop, but they do not replace the frontier across every task. That is the opening: build infrastructure that converts the recurring slice of frontier work into capability the operator can retain, run within a custody domain they control, and keep working after they unplug from the labs.
The corrections, evals, and context that make rented AI useful accumulate behind the vendor boundary. Even when the provider does not train on customer data, the user cannot carry that accumulated capability away. Caletta reverses that flow: route each call to a frontier API or a model inside the operator's custody domain, preserve consented usage as training data and evals, then train a model the operator can retain.
Ownership has a test
Unplug from the frontier labs and reproduce a declared fraction of a real workload, at a declared quality bar, from the operator's own model, corpus, evals, and policy. That fraction is the Sovereignty Ratio. It is task-scoped and reported with the hardware, evaluator, cost, and held-out period. If it does not rise, this is only better-instrumented dependency.
Capability you own, not capability you rent
The public mlx-go-iroh module provides peer-to-peer transport over go-iroh. The self-hostable training coordinator and Go-native inference runtime for Apple silicon are built and publication pending.
Every link has two settings
The Exchange Dial records what crosses each connection: nothing, metrics, weight updates, derived signal, or raw work. The operator also sets what the machine will accept and how incoming contributions are checked. Weight updates are not equivalent to “nothing leaves”; adapters can reveal training data, so the privacy claim changes with the selected level.
Compute no one else can switch off
A custody domain can be one laptop, a workstation, an on-prem cluster, or datacenter capacity under the operator's administration. The first measured cross-machine step used two Macs: they independently fine-tuned, exchanged, and merged one adapter to identical bytes across three rounds, documented with its failures in Two Macs, one adapter. That proves adapter exchange and merge on two machines. Coordinated group training, privacy against update leakage, and progress after node loss remain ahead. Where this work points long-term is written out in Democratized distributed training.
Runs leave evidence
Runs emit network rows, logs, traces, and reproducible receipts. A receipt can bind declared inputs, outputs, and evaluations; it is not hardware attestation and cannot prove that no hidden step occurred. The receipt note records that boundary.
On your side of the line
Private by default is a boundary, not a setting. The security layer, policy, brokered secrets, routed egress, and operator-owned evidence, exists so the loop runs on your side of the trust boundary and stays there.
The Caletta-controlled path is Go-first and designed to keep working without a vendor API after the exit test passes. The longest-standing public proof a reader can check is langchaingo, the Go LLM framework, alongside the published transport substrate. Other pieces are built but publication-pending and are named as such. Nothing here links to a repository a reader cannot clone today.