bound is a deterministic control harness placed after a meaningful execution step and before the agent chooses its next action.
The agent still owns planning, code generation, tool use, and strategy selection. BOUND only determines what the control loop should do next.
In an end-to-end ChatGPT experiment, the first implementation passed one of three required checks. BOUND returned REPLAN. The agent selected a materially different strategy, after which all three checks passed. BOUND returned ACCEPT, and the agent stopped refining that objective.
I also dogfooded bound on its own release. The repository contains the real run trace and generated integration report: 513 full-suite tests passed, 34 targeted tests passed, no unexpected artifacts were observed, and unavailable telemetry remained explicitly null.
This is experimental. The control loop works, but I have not yet demonstrated that it reduces unnecessary work or improves outcomes across workloads. A paired baseline-versus-bound benchmark is next.
I’d especially appreciate feedback on the control abstraction, contract overhead, and what evidence a convincing benchmark should capture.