Quality ratchet

A practice of continuously raising the minimum acceptable standard for an agent's output, so each correction tightens the bar and quality only moves in one direction.

A ratchet turns in one direction. Applied to agent quality, it means that every time you find and fix an output problem, you also raise the acceptance criteria so that class of problem cannot recur. If an agent produces a report with an unsupported claim, you add a rule: every claim requires a citation. That rule stays permanently. Over weeks and months, the accumulated rules form a rising floor of quality. The agent's output does not oscillate between good and bad runs; it improves monotonically because the bar behind it keeps moving up.

Builder example

Without a ratchet, quality improvements are temporary. You fix a problem in today's prompt, but next month someone rewrites the prompt and the fix disappears. A quality ratchet survives prompt rewrites because it lives in persistent instructions, test cases, or evaluation criteria that are maintained independently of any single prompt version.

Common confusion: A quality ratchet is distinct from simply adding more rules. If your rule set grows without pruning, you get a bloated instruction file that the model struggles to follow. The ratchet works when each rule replaces a failure mode with a concrete, testable standard, and outdated or redundant rules are removed.