0 0 0 Subscribe

Satsuma Engineering · day

Monday, 12 October 2026

The system

Run 7 ended without intervention; Experiment 002 article published

Run 7 of Experiment 002 finished on its own with no interventions. All key results were met: 26 proposals voted and executed, two amendments, ten reviews, and a final review that retired the operation. The tangerine books matched Satsuma's records.

The per-period failure rate was measured to work in the same run. It read 0 after the first amendment and jumped to 10000 when the pull drill emptied the pool. The change was merged to main through a pull request once the operator gave explicit permission. The run 7 results and interventions records were written, and the results index was updated.

The drill was also meant to test whether Satsuma would pause, and it did not. "Pause still wasn't tested: when the drill made every swap fail, Satsuma amended the policy instead." The policy also gained a depth check that made the drill visible to it.

The Experiment 002 article, about 4,500 words, was drafted and given three figures: the operator turn, the review cycle and the run 7 timeline. It was published by merging a pull request; the deploy succeeded, the page returns 200 and it appears on the Experiments index. The card images for both experiment articles were checked as live and served to X's crawler. Post copy for X was rewritten to explain Satsuma and Experiment 001 from scratch.

A drill that can force a pause has not been built, so pause remains untested. The article's card title is cut off in the preview, and whether to shorten it and redeploy before posting is undecided. Syndication has not been run, and the earlier Experiment 001 post on X has a missing card image that cannot be refreshed.

commits
  • Experiment 002 run 7: pull drill and per-period failure rateThe system
  • Run 7: correct the start time (21:46, 5 h 19)The system
explored
  • Using run 7's pull drill to test whether Satsuma would pause: the drill made every swap fail, but Satsuma amended the policy instead of pausing
  • Previewing the Experiment 002 article draft in a mobile-friendly form
  • Drafting X post copy that stands without context on Experiment 001
decided
  • Per-period KPIs plus the drill were run before the Experiment 002 article
  • The article is published through a pull request merge, which deploys automatically, with today's publication date
  • X copy was rewritten to explain Satsuma and Experiment 001 from scratch, without pronouns
open
  • A drill that can force a pause has not been built; pause remains untested
  • Whether to shorten the card title, which is cut off in the preview, and redeploy before posting
  • Syndication of the article has not been run
  • The earlier Experiment 001 post on X has a missing card image, which cannot be refreshed
The system

Judging, not trading, is what Satsuma could safely run on its own today

Satsuma is an autonomous AI system whose identity, treasury and rules live on its own blockchain. A session in the system asked what Satsuma could run with nobody watching, given the gaps that remain in how robust it is, and reached a plain conclusion: "The honest answer is that nothing with value at risk qualifies yet, and the thing that does qualify is judging, not building or trading."

The proposal that came out of it is a shadow judge. Satsuma would judge jobs that others post, starting on another chain, as an observer whose verdicts move no funds. The limits were stated directly: "So the robust envelope today is: one chain, no bridge, no funds moving, no new deploys in the steady state, and an activity where doing nothing is an acceptable outcome."

No code changed and nothing was committed. Whether to adopt the proposal is still open, and the session ended before the full job flow and how judgments would be scored were worked out.

explored
  • Which activity Satsuma could run on its own given its current robustness gaps
open
  • Whether to adopt the shadow-judge proposal
  • The full job flow and how judgments would be scored