# Study 2 preflight and execution

Run these commands only from a clean committed research worktree. No GUI model control is part of
the protocol.

```bash
PYTHONPATH=src MPLCONFIGDIR=/tmp/agent-harness-mpl .venv/bin/python -m pytest -q
PYTHONPATH=src .venv/bin/python -m agent_harness.cli validate
PYTHONPATH=src .venv/bin/python scripts/audit_study2_design.py
PYTHONPATH=src .venv/bin/python scripts/preflight_study2.py
```

The outcome-blind preflight uses `lms server start/status/stop` for lifecycle and LM Studio's native
REST API for exclusive residency. It verifies pinned repository heads, dependency versions, disk
headroom, exact agent model variants/contexts/reasoning defaults, tokenizer hashes, custom tool
calling, and embedding dimensions/normalization. It writes
`results/reports/study2_preflight.json` and unloads all models before stopping the server.

The main and reliability executions are resumable:

```bash
PYTHONPATH=src MPLCONFIGDIR=/tmp/agent-harness-mpl .venv/bin/python -m agent_harness.cli run-study2
PYTHONPATH=src MPLCONFIGDIR=/tmp/agent-harness-mpl .venv/bin/python -m agent_harness.cli run-study2-reliability
```

The main command must yield exactly 840 completed deterministic cells. The reliability command
must yield exactly 72 stochastic sensitivity cells. Any incomplete infrastructure attempt is
retained under `results/infrastructure_attempts/`; it is not a scored model outcome.

## Completion record

The outcome-blind preflight passed. Execution revision
`58933d6fa8af09fcc5a832fb3b523ffb4182bc50` produced exactly 840 main and 72
reliability cells with no missing identity and no scored infrastructure
failure. Both run reports end with an empty model-residency set and a stopped
LM Studio server. The deterministic analysis at revision
`bab257bfc30d9959e31ffdd2ba2c9ebf2c5e575b` is under
`results/derived/study2`.