Loop 023
The self-improving champion loop
A bounded optimization loop that tests targeted challengers, protects an independently evaluated champion, and rejects suspiciously easy wins.
Ready-to-use prompt
Copy the loop
Keep three pieces of state in memory: the champion (the best current genome plus its gate score), a budget starting at [N], and a log of every tried genome and score. Each cycle, if the budget is zero, stop and return the champion. Otherwise, reduce the budget by one, read the latest failure in the log, and propose one targeted change to the champion that addresses it. Skip any change already tried. Score the challenger on a working signal: a cheap measure you may tune against. If it is not better than the champion, log it and continue. If it is better, freeze the challenger and run the gate on fresh examples you did not inspect while editing, plus the guard checks in [safety]. Accept the challenger only if its gate score beats the champion by [minimum margin] and no guard regresses; otherwise keep the champion. Log the attempt and repeat. Keep the working signal and gate separate. Treat a suspiciously easy win as Goodhart's law in action and reject it. If you are uncertain, keep the champion.
Verify / stop
The budget is exhausted and the best verified champion is returned.
Every challenger is logged, the final champion has the strongest accepted gate score, and no accepted change regresses a guard check.
Context and guidance When to use it, steps, safety notes, and related loops
Use this when
Use this to improve a prompt, policy, configuration, or other testable artifact when cheap iteration is useful but final acceptance must use fresh evidence.
How to run it
- Initialize the champion, independent working and gate signals, guard checks, improvement margin, budget, and experiment log.
- Use the latest failure to propose one untried targeted challenger, then score it on the cheap working signal.
- Freeze promising challengers and evaluate them on fresh gate examples plus every required guard check.
- Promote only a meaningful, regression-free gate win; log every outcome and return the champion when the budget reaches zero.
Why it works
Separating the tunable working signal from a fresh acceptance gate limits overfitting. The champion rule makes regression the default-safe outcome, while a fixed budget keeps the search bounded.
Implementation note
Define [N], [minimum margin], the working signal, gate examples, and [safety] before starting. Do not reuse gate examples for editing or silently weaken a guard after a failed challenger.