Autonomy level from eval results and blast radius

Before an agent task type may act on its own, the platform team sets its autonomy level from two facts: how it did on the offline eval suite (passes out of runs) and its blast radius ("low", "medium" or "high": how much damage one bad action can do).

Write autonomy_level(passes, runs, blast_radius). It returns one of four levels, from most to least freedom:

• "auto": the agent acts on its own.
• "approve": the agent acts only after a human approves each action.
• "suggest": the agent drafts the action and a human carries it out.
• "off": the task type does not run.

Step 1. The pass rate passes / runs picks a level from this table. Thresholds are inclusive: a pass rate exactly on a threshold earns that level. Below the suggest threshold the level is "off".

Step 2. A small eval proves little, so the run count caps the level. With fewer than 100 runs the level is at most "approve"; with fewer than 20 runs it is at most "suggest". A cap only ever lowers a level, never raises it.

Fail closed. Zero runs, or a blast radius that is not exactly one of the three strings (case matters), gives "off".

You can rely on 0 <= passes <= runs. Compare without floating-point rounding, for example passes * 100 >= runs * 95.

Examples:

• autonomy_level(196, 200, "low") returns "auto": 98% clears 95%, and 200 runs is enough evidence.
• autonomy_level(196, 200, "high") returns "approve": a high blast radius never earns auto.
• autonomy_level(19, 19, "low") returns "suggest": 100% would be auto, but 19 runs caps it at suggest.