- Approved image
- Candidate image
Waiting for execution
The runner will retain incomplete evidence rather than infer a pass.
Measured comparison
| Metric | Approved | Candidate |
|---|
Timing is workload-reported, excludes container startup, and uses medians across trials. Time per sample is amortized processing time, not request latency. Memory is process RSS on CPU or allocated tensor memory on CUDA. This is a small qualification suite, not an LLM serving benchmark.
Release policy checks
Inspect output cases
Errors are calculated by the controller against frozen reference values. Rows show the worst observed error for each candidate case.
| Case | Max absolute error | Tolerance | Evidence |
|---|
Execution trail
Versions, policy and input identity
Logs and evidence files
Exports include workload code, fixtures and outputs. Keep private bundles inside your approved environment.