gauntlet-loop
Gauntlet Loop
Give a lead agent the destination and bar. Let it choose the route. Keep building and judging separate.
1. Set the bar
State the goal as an observable outcome while preserving the user's constraints. Use a supplied reference when it is inspectable. Otherwise propose the strongest concrete comparison or measurement available. For a quantitative outcome, state a specific value at a stated scale. For a qualitative outcome, state the reference or the observable criteria. Explain its relevance in one sentence. Ask the user to confirm or amend it before round 1. A placeholder that the user must fill later is not a bar.
Define how a critic can compare the real artifact with the bar:
| Artifact | Useful bar |
|---|---|
| Visual product | Rendered reference, screenshots, or interaction recording |
| Software system | Passing behavior, performance, recovery, or security checks |
| Writing | Reference passages plus factual and structural checks |
| Research | Source quality, coverage criteria, and reproducible calculations |
For a visual bar, fix the capture protocol alongside the reference — viewport, theme, seed data, animation state — so both sides of every comparison are captured the same way. A comparison whose two sides were captured differently returns UNJUDGEABLE, and the loop then repairs the inspection path instead of the artifact.