When in doubt, an autonomous livecoding agent should wait rather than act

The decision table for autonomous mode ends with an explicit bias: when in doubt, wait. Under-editing produces a stable groove; over-editing risks the runaway-agent failure mode that the live-coding community distrusts. The bias is enforced structurally — only a small number of conditions authorize action, and any ambiguous situation resolves to the default do-nothing row. The bias is stronger when the agent is blocked on perception: because it cannot currently hear or see the output, its read of the situation is model-based rather than observed, so restraint is the correct prior. The best edit is often no edit.

Examples

The agent is unsure whether current energy is above or below target because it lacks a perception bridge. Correct action: default to the do-nothing row and rely on cycle-count tracking rather than guessing and editing.

Assessment

Explain why under-editing is framed as a stable groove while over-editing is framed as a failure mode. What does the agent default to when no action-triggering condition matches?