Build log
Every entry is a recovery, a substitution or a verification.
Never a capability demo. A robot ran 100 metres in 8.64 seconds this month — nobody needs another video of one walking. What nobody has shown is a machine noticing it was wrong and doing something about it.
A hotel-standard room turnover dispatched across a robot arm, a mobile rover and a humanoid. The humanoid took only the bed — the one task no other body could do. The fixed arm correctly took nothing. In simulation.
"Make the bed" expanded into locate sheets → identify orientation → grasp corner → secure corner → verify, but only because no single skill could do it whole. Where one could, it did not expand.
An inspection route is obstructed mid-job. ReasoningOne diagnoses the failure in terms that generalise across a fleet, replans, and escalates to a person when another attempt is the wrong answer.
Identical runJob() call, unchanged, driving two physical machines of different embodiment from different manufacturers.
Confidence drops partway through a task; the router moves to a different model and the job completes. The routing trace shows both hops and what each cost.
A real facility, a real invoice, and the first verified autonomous job that belongs to somebody else.
Planned entries are shown as planned. A log that quietly backdates its promises is worth less than no log, because the people reading it are the people deciding whether to let us onto their site.