Machine Log, week of 20 April, part two: the machine scores its own homework
The weekend brought a contribution loop, the first deployment to a rented server, and a research run that scored itself 33, then 100, then stopped.
This entry covers the second half of the week of 20 April 2026 — Friday to Sunday; it was written and published on 4 July 2026 as part of the Machine Log catch-up series.
Friday 24 April, eleven changes: the contribution loop. The route by which a supporter’s idea arrives, waits for review, and either becomes work or gets declined was split into clear stages, so an idea can no longer fall into an untracked gap between arriving and being looked at.
Saturday 25 April, eleven more: the automation that runs missions was hardened, runs were made checkpointable — stoppable and resumable without losing work — and tooling was built and tested for publishing evidence packs to a public network. Built and tested is the full claim: no pack was published to any public network that day or since, and this log will say so plainly whenever the subject comes up. Saturday also brought the machine’s first deployment to a rented server, moving it off the operator’s own hardware for the first time. Sunday 26 April added nine changes: an upgrade to the operator’s control panel, mobile fixes, and a written note on open-source stewardship.
Then the machine did the week’s most important thing: it graded itself. On Sunday afternoon, two automated research runs executed against the growth-index mission — a bounded question about how Britain’s growth is measured. The first run scored 33 out of 100 on the mission’s own benchmark. The second scored 100, with ten cited sources and a 2,218-word evidence pack, and was kept. A third pass produced no improvement and was discarded. Keep what scores, kill what doesn’t, log all three — that is the entire operating philosophy in one afternoon.
The state offered both halves of the same lesson that week. On Tuesday the government published its Reformed National Pricing delivery plan for the electricity system, with dated decision points from the second half of 2026 (GOV.UK, 21 April 2026). On Friday, reporting revealed DSIT had mothballed several of its own internal AI pilots to refocus on legacy systems, redirecting effort to speeding planning decisions (PublicTechnology, 24 April 2026). A government willing to close its own AI experiments is applying the discard rule too. It is less embarrassing to kill a run that scores 33 than to publish it.
About forty changes, one deployment, one benchmark, one discarded run. Next: the week Future Union launched — and the five briefs that never made it out.