Goal
Run LoopSpine explicitly on ten real repository tasks before any global promotion. Synthetic benchmarks and the bundled demo do not count.
Task register
For each task, preserve the exact proof, elapsed time to proof, Sawyer intervention count, and whether LoopSpine stopped incorrectly. Record the result with npm run dogfood:record -- completed-run.json.
Four metrics
- Verified completion rate
- Sawyer intervention rate
- Median time to proof
- Incorrect-stop rate
Promotion gate
Do not change global skills during the pilot. Consider narrow entry-skill promotion only after 10/10 real tasks, at least 80% verified completion, at most 20% intervention, and reviewed incorrect-stop residue.
Goal
Run LoopSpine explicitly on ten real repository tasks before any global promotion. Synthetic benchmarks and the bundled demo do not count.
Task register
For each task, preserve the exact proof, elapsed time to proof, Sawyer intervention count, and whether LoopSpine stopped incorrectly. Record the result with
npm run dogfood:record -- completed-run.json.Four metrics
Promotion gate
Do not change global skills during the pilot. Consider narrow entry-skill promotion only after 10/10 real tasks, at least 80% verified completion, at most 20% intervention, and reviewed incorrect-stop residue.