The lab grows up

August and September were the months the lab stopped being a set of scripts and became something that runs itself.

Marilyn v5 landed. A 931M-parameter trunk on about 40B tokens, recovered from a stall through its workflow rather than by hand, with two more variants built beside it and all three then parked on purpose. See the v5 programme.
The fleet got an operating system. A shared Go library for Temporal workers, tiny single-purpose workers built on it, a nightly backup that proves itself by restore, a weekly disaster-recovery drill on a disposable cloud box, and break-glass recovery on two continents for about six dollars a month. See the operating stack.
LiLM’s ingestion factory was built in one autonomous week, gated at every stage, and passed a real-corpus test that also told me its summariser was truncating and its knowledge graph was nearly empty. Both findings are on its losses page.
The question changed shape. Every way Marilyn learns ends in a training run. LiLM asks whether there is another way, and until it answers, the GPU time waits.
This site was reviewed before publishing by five separate AI sessions, each given a reader’s persona and no access to anything but the pages: a hiring manager, an enthusiast, an attacker, and two less charitable readers. The hardware inventory came off the site as a result. Open weaknesses that were still live came off for the same reason; the ones that were fixed are all still here, on the losses pages, with the fix.

What changed: the lab now has the discipline of the day job, and the day job has the lab’s habit of measuring before believing.