Monday, 2 November 2026

Satsuma's operating loop, the part of the autonomous system that watches the products it runs, assesses incidents and sends defects to be fixed, was put through a live run to test four of the judgment checks its operating loop depends on, with a cap of 6 hours. The run finished in 3 hours of loop time. The incident assessor passed: a recovered outage was closed as External, and a planted defect was filed and sent to a fix. The fix reviewer passed: three fixes approved, no refusals, and all three landed and passed their watch windows.

The backlog review check came back partial, and the cause was a bug in the loop: reviews kept reading the originally deployed code after a fix had landed. "The cause was a lookup order that always preferred the originally adopted repo over any later fix copy, and the backlog review was never given the code at all." The fix was merged in the satsuma and satsuma-experiments repositories; 1,302 tests passed locally and CI passed. With the budget for hosted CI runs at $0, the decision was to run the full suite locally and merge only where branch protection allowed it without CI.

The fix was proven twice. A scripted review-after-fix scenario passed in 6.6 minutes with no live model calls. Then a capped live check of 10 minutes on Opus showed an already-fixed defect was not filed again, and the backlog review read the latest fix code. With that, the protocol checklist is complete: every capability passed either in a scenario or in the live run.

The run needed three restarts of its drills and one of the loop. A stale idle watcher left over from an earlier run was stopped by its process ID and noted in the interventions log. The verdict of the fourth judgment check is not visible in the record.

GitHub sets a spending budget for each paid extra on an account. GitHub Actions, the service that runs automated tests when code changes, is one of them. A check was made to set this budget to $0. Nothing needed changing: the GitHub Actions budget is already $0, with "stop at limit" on. A task scheduled in a desktop app had already reset it, and the Codespaces, Packages and Git LFS budgets are at $0 too.

One decision came out of the check: if Satsuma's live updates need a code change while the budget stays at $0, the tests run locally and the update reports that CI did not run. Still open: CI on the private repositories may not start while the budget is $0.

A live run started to get verdicts on four judgment checks of Satsuma's language model: assessing an incident, reviewing a fix, reviewing pending work, and a redesigned adoption probe. Every run so far has used Claude Opus, and it answers only the judgment steps while the mechanics stay scripted; the run caps at 6 hours and stops once all four checks have a verdict, with its records to be archived without a write-up.

Whether all four checks reach a verdict is still open, and a full 6-hour run needs the Mac it runs on plugged in.