About this record
A useful record of learning by doing.
This site follows a local AI lab as its hardware, models, runtime choices, and operating practices evolve. It shares enough technical detail to help someone build and measure their own setup.
What belongs here
Hardware configurations, runtime settings, reusable agent workflows, measured results, failed experiments, and the reasoning behind changes. The journal keeps reversals and superseded choices because the path to a working configuration is part of the evidence.
Three kinds of evidence
- Verified configuration
- Inspected hardware, running arguments and routing. It establishes what is configured on the inspection date, not performance.
- Measured experiment
- A dated workload and protocol. Scores retain units, task counts, runtime and limitations. Catalog text and configured capacity remain separate.
- Reusable guidance
- Lessons from the lab and links to primary documentation. A reference template is identified as a template, and untested settings are not presented as qualified.
The publication boundary
The public record excludes private project identities, business work, personal information, credentials, network addresses, user paths, and original private prompts. It contains curated technical facts and reusable examples. The site serves committed content independently of the lab machines.
How it stays honest
Both GB10 containers, installed coder settings and coordinator selectors were reviewed on Oct 23, 2026. Other hardware and Studio launcher observations retain their individual dates. The matched DeepSeek/GLM panel, short serving probes, cache replay and memory observation have separate protocols. The October 1 idle snapshot remains historical. Earlier experiments keep their own dates; the September Studio study retains its chat-adapted HumanEval and custom-suite caveats. When a direct query conflicts with a dashboard label, the direct evidence is recorded.