Autolab ships confidence-first autoresearch CLI stack

The sharp insight here is that bad keeps are costlier than clean discards, and this toolkit operationalizes that idea into a reproducible loop.

• –`autojudge` reframes tiny metric deltas as statistical confidence decisions, which is exactly what noisy overnight autoresearch runs usually lack.
• –`autosteer` adds lightweight portfolio logic (explore vs exploit) that can reduce random-walk experimentation without claiming causal certainty.
• –`autoevolve` is the highest-upside piece for teams, but it also introduces the most orchestration complexity around branches, compute, and merge hygiene.
• –The author’s own caveats matter: confidence estimates need enough recent runs to stabilize, and early-stage tools can overfit to local experiment dynamics.