Generated Code Quality Gates

Teams seeing AI-generated diff volume outrun reviewer capacity can build a narrow auto-review path for low-risk changes. RADAR gives a concrete pattern: classify diffs by source, exclude sensitive files and scopes, score each diff with a risk model, run LLM review, run deterministic checks, and auto-land only within configured thresholds. Meta reports more than 535K RADAR-reviewed diffs and more than 331K landed diffs, with lower revert and production-incident rates than non-RADAR diffs in the deployment. A small version can begin with codemods, generated runbook changes, or allowlisted owners, then compare revert rate, incident rate, and review wall time against ordinary review.

Developers often receive generated code before they have a trusted test suite. TRAILS points to a buildable check for that gap: extract behavior categories and preconditions from the natural-language spec, generate candidate inputs, execute the candidate program, and ask an LLM to judge each input-output pair against the spec without seeing the code. The method improved Matthew correlation coefficient over zero-shot chain-of-thought baselines on LiveCodeBench and CoCoClaNeL, though it used more tokens per task. This fits code-generation surfaces where accepting a wrong function is costly and a full test suite is unavailable, such as internal scripting tools, data transforms, and generated helper functions.

Security teams running agentic vulnerability repair can store repair attempts as reusable experience, keyed by CVE or CWE. EvoRepair records vulnerability analysis, repair strategy, trajectory analysis, reusable rules, and follow-up notes; it retrieves related experience before patching in Docker, then scores and updates the store after each attempt. The paper reports 93.47% on PATCHEVAL and 87.00% on SEC-bench with GPT-5-mini, ahead of 12 automated vulnerability repair baselines in its comparison. A practical pilot can start with one recurring weakness class, such as path traversal or command injection, and measure whether the agent repeats fewer failed edits and reaches passing security tests in fewer turns.