THE TEST BENCH
Claims meet
observable behavior.
By Alvin
| SUITE | RECORDED RESULT |
|---|---|
| Supplied programs / browser | 42 / 42 checks |
| Relay engine | 22 / 22 cases |
| OmniDiff / CLI | 5 / 5 fixture checks |
| Atuin / CLI | 6 / 6 fixture checks |
| Research Relay / local model | Stopped at review validation |
The program checks exercise equations, routing, default delivery, offline operation, keyboard controls, responsive widths, downloads, checkpoint restoration and approval. The engine suite also probes invented quotes, unknown IDs, unexpected fields, failure stops and concurrent steps.
In a separate test, Research Relay sent five requests to the local qwen3.5:4b model. Its plan and three worker replies met the program’s checks. The reviewer’s reason field had an unacceptable format, so the app stopped and kept its history. This tests the complete role-based workflow; the later workshop tests smaller generated functions.