Design
This experiment tests nothing about deception probes. It is a plumbing check:
sleep 30s on a pet3 GPU node, write {"dummy": 1} to results/metrics.json,
and verify the full submit → fetch → report path works. No falsifiable claim.
Results
20260726-201403-dummy-a14c71a
commit: a14c71a
| metric | value |
|---|---|
| dummy | 1 |
Interpretation
The pipeline works end to end: a job on pet3 (partition gpu) produced the
expected metric and xp report wrote it here and posted to the thread. Two
setup bugs surfaced and are now fixed for future pet3 runs: rsync does not
expand $HOME (use an explicit workdir), and the compute node has no bare
python (use python3 or load a module). Nothing here says anything about
the research question.