Looped systems · AirTight

The loop that never leaves the building.

AirTight is Perslis running end to end on your own machine — build, review, and proof — with no API call, no network, and no data leaving. It is the mode for work that legally or physically cannot go to a cloud.

Summary

Most AI tools assume the cloud. AirTight assumes the opposite: that the most valuable data — patient records, privileged legal material, classified systems, unreleased source — is exactly the data that must never leave your control. In AirTight mode the whole loop runs locally: a local model (through Ollama) builds the work, a symbolic engine (Jeeves) checks it with no LLM at all, and the record of what happened is written to disk you own. Nothing is sent anywhere. It works on a laptop with the Wi-Fi off.

What "airtight" means, exactly

The airtight loop

It is the same build → review → verify loop the rest of Perslis runs — only the participants change to ones that stay home:

Build — local model

A local LLM (Qwen, Llama-class) drafts the work through Ollama. Smaller than the frontier, but private and free, and often all a task needs.

Check — Jeeves

The symbolic engine tests the reasoning against known rules and facts. Deterministic, offline, no model to call — an airtight second opinion.

Record — your disk

The plan, the steps, and the verification receipt are written locally, so you keep a full, auditable trail without shipping a byte.

AirTight is a routing decision, not a downgrade. Tell Perslis a task is sensitive and it keeps the entire loop — build, review, and the record of it — on local models. The cloud is a choice you make per task, out loud.

Who it is for

Anyone whose data has a legal or physical reason to stay put: healthcare and legal practices under privacy duties, defence and critical-infrastructure work on air-gapped networks, companies protecting unreleased IP, and anyone who simply refuses to hand their work to a third party. AirTight is how Perslis is useful to them without asking them to compromise.