# Local release measurements

Repository tools provide fixed mixed-value and full-Enron workloads:

```text
poetry run python tools/benchmarks.py components --output COMPONENT_REPORT.json
poetry run python tools/benchmarks.py pipeline --area NEW_TRIAL_DIRECTORY --output PIPELINE_REPORT.json
```

The component workload checks strict/logical normalization and CPU/CUDA reference
parity, first/warm calls, and isolated warm GPU allocation/transfer/kernel timings.
The pipeline workload scans all 517,401 pinned Enron records with three local
workers, publishes a snapshot, reopens it and runs report/export operations.
`tools/measure_windows_job.ps1` measures elapsed time and aggregate committed host
memory across the process tree, records code/dependency/hardware provenance, and
rejects source changes during a trial. Its memory metric excludes GPU VRAM.
`tools/release_limits.psm1` freezes limits from three trials of each workload and
evaluates independent holdouts against checksummed limits. Failed limits retain
failure receipts and fail the command. Raw workload reports establish parity and
measurements; release acceptance also requires frozen-limit and quality checks.

`poetry run python -m pytest tests/gates` proves the formatting, typing, coverage,
dependency and broken-runtime-requirement gates actually reject controlled faults.