loop-engineeringlisted

# Loop Engineering ## Overview Use a bounded control loop where each iteration reduces uncertainty or advances an observable result. ## Entry Contract Before acting, establish: - one objective and its boundaries; - observable completion evidence; - safety, authority, and compatibility constraints; - an iteration, time, or tool budget; - the smallest verification command or observation that can test the next action. Infer only reversible, low-risk defaults. If missing information would materially change public behavior, cost, scope, or an irreversible action, return `needs-user-decision`. Classify the task before acting: - **trivial:** run `goal → action → verification` once; - **local:** run the full loop with local evidence; - **cross-cutting:** run the full loop and apply Graph Escalation. Read [loop-protocol.md](references/loop-protocol.md) before iterating and [stop-conditions.md](references/stop-conditions.md) before declaring an outcome. ## Run the Loop Repeat `plan → act → verify → decide`: 1. **Plan:** choose one falsifiable hypothesis, one bounded action, and its expected evidence. 2. **Act:** perform the smallest authorized action that tests the hypothesis or advances the objective. 3. **Verify:** run the planned check and compare expected with actual evidence. 4. **Decide:** complete, retry with new evidence, replan, block, or request user direction. Track the protocol state in task memory or temporary scratch space. Never add loop transcripts, chain-o