source/HANDOVER.md

# OpenSysOne handover

**Recorded state, 17 September 2026, 09:22 UTC:** training, final validation,
evaluation and profiling completed with exit 0. Training remains stopped; both
Spark GPUs were idle at completion. Preserve all selected and resumable artifacts.

The selected Qwen3-4B decision scorer retains Spark B step-1,500 weights, unchanged
at expanded branch step 0. It achieved **92.90%** accuracy on the original test
and **72.92%** on the Social IQA family holdout. The measured implementation is
slower than both pretrained inference baselines. See [RESULTS.md](RESULTS.md).

The recorded remaining services are the calibrated loopback API on **18081** and
the browser playground on **7466**. Process IDs, exact source revisions, checkpoint
hashes, run paths and inspect/stop/restart commands are preserved in the
[full operational handover](docs/operations/handover.md). Check those records and
the live process identity before acting; this document does not restart any job.

For a continuation, read [PLAN.md](PLAN.md) and [RESULTS.md](RESULTS.md), then the
[full handover](docs/operations/handover.md) and relevant infrastructure page under
`/home/andy/ai/opensysone/gx10-reference`. Keep the 16 GiB per-process allocation
cap and existing model/data/source provenance checks. Historical launch plans in
the archived documents are superseded by the completed state above.

- [Documentation index](docs/README.md)
- [Playground usage and controls](docs/usage/playground.md)
- [Fleet history and serving-pair restoration](docs/operations/fleet.md)
- [Hugging Face publication records and controls](docs/operations/huggingface.md)
- [Complete report, tables and charts](results/report.md)