How it works

One loop, five ledgers, and what it does not do yet.

1. Find work

Metric triggers and failure detectors file issues that carry a code task and a closing probe.

2. Hand it to an agent

A work-claim ledger gives each file to one agent at a time: an open-model code loop, Copilot or Claude Code.

3. Prove it live

Probes declared per service run after each deploy. A regression reverts the push and files its own issue.

4. Score the decision

Each decision logs a prediction before it acts; the outcome is scored later and feeds a trust score per kind of action.

Why the name

W. Ross Ashby's ultrastable system (1952) changes its own settings when an essential variable leaves its limits, so it stays viable without an operator. The trust ledger, doctrine amendments and probe-and-revert loops do that for software.

Honest gaps

One installation and one customer. There is no installer and no second tenant yet.
The core still shares a database with QNFO's research applications; separating it is the first pilot milestone.
The prediction match rate is scored by the system itself. An outside audit is a pilot goal.
A buyer needs a short default policy pack; today the policy is long and written as owner directives.

Join the pilot

Ultrastable is an early-stage product. Every number on this site is read from a live system's own records. See the checks · Product · Team · FAQ · Contact · llms.txt