Validate & Benchmark
Use the smallest test that can actually prove the claim:
| Claim | Minimum meaningful evidence |
|---|---|
| Parser/config behavior | Focused unit or contract test |
| Exact checkpoint builds | Successful build with exact model revision/config |
| Task correctness | Model-owned E2E input, oracle, comparator, and thresholds |
| Target compatibility | Full build/load/run on that target and software cohort |
| Performance | Fixed workload, warmup, repeated measurement, timing boundary, and quality gate |
Start repository validation with ownership checks:
python3 tools/model_ci.py validate
Then run the exact E2E manifest or focused test named by the model change. A skipped GPU preflight is not a pass, and a documentation build is not model parity evidence.
For performance, record model/revision, bundle checksum/config, execution path, hardware/software cohort, complete command, input, warmup, measured iterations, metric boundary, result, and task-quality result.
Use Testing Reference and Benchmarking Reference for command details. The Validation and Benchmarking Tutorial is the course-style lab.