# Working on OpenSysOne

Read HANDOVER.md, PLAN.md and RESULTS.md before a continuation. This project uses
real shared GB10 machines. The infrastructure specification is the GX10 repository:
on the Mac `/Users/andy/code/gx10`, on GX10 see the handover's documentation copy.
Read the relevant operational page before changing a machine.

GX10 currently has no ConnectX connection. On 2026-09-16 the user assigned all
three machines to this task and explicitly authorized stopping their workloads.
The Spark serving pair has been stopped and its restart commands recorded in
[docs/operations/fleet.md](docs/operations/fleet.md). Treat the hosts as separate
memory pools. Use `free -b` and GPU process inspection before loading a model;
reclaim GX10's inherited SSH OOM
exemption for training jobs. Keep this smoke's 16 GiB allocation cap.

Source and small results belong here. Model/checkpoint files belong under
`~/ai/models/opensysone` and `~/ai/opensysone/runs`; never commit secrets or weights.
Use GX10's existing Python read-only for the smoke. New dependencies go in an
isolated environment. No service or network changes are needed to run this code.

Maintain model/data/source provenance and checkpoint before evaluation.
Synthetic smoke results prove wiring, not task generalization or calibration.
At handover, record source revision, run paths, exit status, and how to
inspect/stop/resume any job left running.