A deployment looked suspicious until older-version jobs told a different story. Explain decisions using the clues and authority available when people made them.
A job ID helps you find one export. Put it on every metric and it can create a new storage problem. Keep the count and the individual story in useful places.
A tiny model predicts zero where the answer should be three. Follow one training step, save the learned values and see why changing the input is a different job from changing the model.
A retention condition may sit unused for many turns, then decide what can be deleted. Follow it through a smaller working context before counting the tokens saved.
A copied runbook and a shell command deserve different decisions. Follow the workdir, tool overrides and saved settings that determine which approval prompt appears.
Selecting fewer experts can save computation without shrinking the whole model to its active size. Follow loading, tool use and the complete local task.
Remembering two maintenance times is not enough. A useful resumed task applies the correction, clarifies its target and checks what happened before retrying.
A calculation can finish quickly while the next waits for memory. Maia 300 joins compute, memory and networking; service trials must show what that changes for users.
Rack designs, controller code and future optics offer different kinds of access. Follow what each contribution lets a builder or operator inspect today.