Evaluation Loops
4 workflows
The keep-or-revert eval loop
ACTION
Improve [target: a prompt, code path, or model] against [a single metric]. Freeze the metric and record a baseline, then each round form one hypothesis, make a single change, and run the same evaluation.
CHECK
Compare the new score to the baseline and keep the change only if it is better-or-equal; otherwise revert it immediately (for example git reset) so every accepted change is a real measured win.
STOP
Stop when the iteration budget is exhausted or the metric plateaus, and return the best checkpoint with the experiment log.
The PR-feedback loop
ACTION
Take the review comments on [pull request] and address them one at a time, making the smallest change that resolves each and noting what you changed.
CHECK
After each fix re-run tests and lint and confirm that specific comment is resolved before moving on.
STOP
Stop when every comment is addressed with green checks, or a comment needs a product or architecture decision from the author.
The two-model review loop
ACTION
Review [change] against [quality bar]. Have a model from one provider review it, verify and apply only necessary fixes, then hand the revised version to a model from a different provider.
CHECK
A round only counts when the findings from each reviewer are independently verified against the actual code or spec, not taken on faith.
STOP
Stop when both models approve the same unchanged version, the round limit is hit, or they oscillate on the same disagreement.
The claim-to-evidence loop
ACTION
List every factual claim [artifact] makes. For each, find primary-source evidence and label it proven, weak, or unsupported.
CHECK
Re-verify each weak or unsupported claim against a second independent source before changing anything.
STOP
Stop when no unsupported high-risk claim remains, or return the rest as needing a human decision.