A small AI task. Checks you can inspect.
Outside15 keeps the outdoor break simple. Behind the return screen, a local agent organizes the observation, checks its evidence, and hands control back to the person.
Keep the failures in the story.
| Model | Passed | Run |
|---|---|---|
| llama3.2 | 8 / 8 | Earlier baseline |
| gemma3:1b | 3 / 8 | Preserved failed experiment |
| gemma3:4b | 8 / 8 | Current extractor |
Eight focused English regression cases. Results reflect each run’s model and validation path; they are not a general model benchmark. These runs used different integration paths, so they are development evidence rather than a controlled model comparison.
Reject. Retry. Return for review.
This is a controlled failure injection, followed by actual inference. Mastra made 2 attempts. The workflow saved zero observations.
A broken check stops Pages deployment.
Python validation → Mastra failure tests → browser flows → static build → repository-subpath and offline checks → GitHub Pages deployment.
Workflow implemented and checked with local equivalents. A successful hosted Actions run is still required after pushing to your repository. Live inference is verified locally and excluded from hosted CI.