Evidence studio and reproducibility

Independent checks

The adapter separately checks selected resource uniqueness, count, capability, health, power, latency and failure-domain constraints. It checks that model phases use selected resources, fit memory and satisfy the supplied model objectives. Those checks are separate from a successful JSON parse.

Inspect and tamper

Use the Evidence Studio to open the actual recorded envelope, select individual artifacts and alter one artifact in the tamper exercise. The consistency verifier must reject a changed payload with unchanged hashes. It does not establish an external signature or prove an arbitrary uploaded envelope came from Garmo.

Benchmark the release

The benchmark repeats feasible, diversity-blocked, memory-blocked and power-blocked scenarios. It includes process startup in wall time, keeps raw samples and checks deterministic artifacts. These are local software timings, not GPU or chip performance claims.

python tools/gsp_cli.py benchmark --repeats 5 --output benchmark.json

Compare controlled runs

Compare two to four runs from your organization. Check whether their input digests match before comparing timings. Different inputs can be useful experiments, but they are not an apples-to-apples performance comparison.