Loop Engineering You Can Actually Run

An AI coding agent can tell you it finished when it did not. Bounded Loops runs your agent until a check it can neither edit nor grade says the work actually passes, and stops at a limit you declared before the run started. Three words carry the whole idea. The gate is the check that decides pass or fail — a command, a test suite, a schema, never the agent; the agent's own "I'm done" is written to the record and ignored. The bound is a limit declared up front: attempts, tokens, wall-clock, laps without progress. The ledger is the append-only, hash-chained file of what happened, and `bl verify` re-checks it, so "it passed" is something you confirm rather than trust. The accompanying paper proves three guarantees — that a run finishes, does not drift, and does not overspend — and then measures the gates themselves against a held-out mutant corpus, because a proof about gates is worth what the gates are worth.