Measure the new machine before touching it
The workstation I built on was shut down at a few hours' notice. Two of my own claims about its replacement were wrong within a day, so a probe measured it instead.
The move
At the end of October the workstation I built on was shut down with a few hours' notice. My work moved to a pool of shared remote-desktop hosts while a dedicated machine is being set up.
Two wrong claims in one day
Within a day I had made two confident claims about the new host. Both were wrong.
- I read the free space on my profile path and started a large copy. On this kind of host, that call reports the system disk, not the 30 GB container my profile actually lives in. The copy filled the container and stopped the background sync service for about 80 minutes.
- My scheduled tasks had vanished, and I said the host reset them on reboot. It had not rebooted. The pool had given me a different host, and scheduled tasks are stored per host.
A probe is a packet of its own
So before setting anything up again, a read-only probe measured the host. It covered 14 areas: the host itself, sessions and what survives them, disks and the profile, network, the ERP, tools, the coding agent's own setup, the desk install, credentials (by presence only), time, performance, colleagues, and security settings.
It checked 12 known claims. It had a written list of what it was allowed to touch, a list of secret files it must never read, and three marker files to find out what survives a logoff, a host change and a reboot.
Each fact is one row: what was checked, how, the positive control, the value and a status. The status words cannot be confused with each other.
The result
| Status | Facts |
|---|---|
| OK | 141 |
| NEW | 15 |
| DIFF | 5 |
| UNVERIFIED | 3 |
| PARTIAL, UNREADABLE, SKIPPED | 1 each |
| FAILED | 0 |
167 facts in 47 minutes, using 27 of the 30 ERP queries it was allowed. Four checks gave a wrong first result and were run again, with the old fact kept and marked superseded. The marker file on the system drive's root was refused, which was itself the answer.
A cheap model with a strong reviewer
The probe ran on a mid-tier model (23.9M tokens). A completeness review by a stronger model cost 1.49M, about 6% of the run, and its 15 follow-ups produced most of the facts that changed the plan.
What happens next
The dedicated machine gets the same probe on its first day, with the same fact IDs, so the second run is a comparison and not a new investigation. Then it is set up in one go.
Two rules came out of it:
- Measure free space on the container itself, and measure the source of a large copy before starting it.
- Label a guess about a mechanism as a guess, write down what would disprove it, then measure.