Evaluate accuracy and speed.
Compare configurations, inspect missed detections and review pass or fail results.
Execute the frozen OpenCV pipeline locally.
The search, including its mistakes.
Development evidence only. Test results are opened after the selection is frozen.
Quality versus throughput
Development imagesDashed line: minimum acceptable recall. Rejected candidates may also fail precision or p95.
AWS, measured on real workers.
Recorded cloud benchmarks of the earlier rules-selected configuration.
| Runtime | Instance | OpenCV / NumPy | Batch median | Recall / precision | Compute rate | Gate |
|---|
What this comparison can support
Resource cleanup
Recorded verification, not a live AWS monitor. Infrastructure usage can appear in billing later.
Evaluation report
Configuration, quality, provenance and known limits in one portable bundle.
Frozen configuration
Selected using development measurements before test evaluation.
Test results
Reproducibility
Evidence boundary
Verified event journal
Catch the regression before it ships.
Frozen configurations, randomized timing blocks, and the detections a faster change would lose.
Paired speedup and uncertainty
95% bootstrap intervals across measurement blocks| Configuration | Gate | Recall | Precision | Batch median | Lost / recovered |
|---|
Inspect the evidence
Use it in your build
The command exits 2 when a bounded configuration fails recall, precision or latency. Invalid input exits 3. Reports attach to the GitHub Actions job, including failures.
The included workflow compares configuration changes using the current detector implementation. It does not compare arbitrary source revisions or enforce branch protection.