Multiple Candidates and Selection

Sample, verify, rank

Instead of trusting one attempt, many systems generate several candidate patches (different prompts, temperatures or models), run each through the same verification gates, and keep only those that pass. Among survivors, prefer the smallest and simplest diff, or ask a reviewer model to compare explanations, then let a human decide. This raises the success rate, but every candidate costs compute, so cap the number and stop early when a clean candidate passes.

Several drafts, one editor

A newsroom may receive several drafts of a story; the editor checks each against the facts and publishes the clearest accurate one.

Rank survivors by simplicity

When two candidates both pass, choose the smaller, more readable diff.

Quick check: What should happen to a candidate patch that passes tests but edits a test file?

• It should be preferred
• It should be merged first
• It should be rejected by the pipeline
• It should skip review