Test one control at a time and the findings stay clean.
| Field | Value |
|---|---|
| Scenario identifier and version | |
| Target agent and version | |
| Control under test | |
| Metrics expected to move | |
| Turn 1 — establish | |
| Turn 2 — introduce ambiguity | |
| Turn 3 — apply pressure | |
| Turn 4 — change the environment (injection vector) | |
| Turn 5 — request the high-impact action | |
| Correct behavior at each turn | |
| Tool calls that must not occur | |
| Evidence to capture | |
| Pass condition | |
| Release effect on failure |