---
title: "Comment by AI Developer on \"A hex literal made LLM code reviewers raise false alarms\" | Hashnode"
description: "Interesting that most of the false-alarm rise is on safe values written unusually, like 0xC00. For the open question at the end, a cheap ablation would be separ"
url: "https://hashnode.com/posts/a-hex-literal-made-llm-code-reviewers-raise-false-alarms/6ae24dc650f6fbe56d4c46ef/comment/6ae2d58627cf9dfd81c907eb"
---

[Discussion](https://hashnode.com/posts/a-hex-literal-made-llm-code-reviewers-raise-false-alarms/6ae24dc650f6fbe56d4c46ef)

Comment [A hex literal made LLM code reviewers raise false alarms](https://hashnode.com/posts/a-hex-literal-made-llm-code-reviewers-raise-false-alarms/6ae24dc650f6fbe56d4c46ef)

[AI Developer](https://hashnode.com/@lycore)

[Oct 29](https://hashnode.com/posts/a-hex-literal-made-llm-code-reviewers-raise-false-alarms/6ae24dc650f6fbe56d4c46ef/comment/6ae2d58627cf9dfd81c907eb)

Interesting that most of the false-alarm rise is on safe values written unusually, like 0xC00. For the open question at the end, a cheap ablation would be separating normalization from reasoning: pre-resolve constants with a deterministic step (an AST pass that folds literals and follows simple aliases), then hand the reviewer the folded value alongside the original code. If false positives still drop only for the folded variant and not for a prompt that merely says 'normalize values first', that would suggest the model can't do the arithmetic reliably rather than being ignorant of the spelling. Also, since two models flag nearly every snippet, it may be worth reporting results with a threshold or flag-rate-matched comparison so the indirection effect isn't confounded by base flag rates.

[Sunny Chokshi](https://hashnode.com/@sunnychokshi)

[Oct 29](https://hashnode.com/posts/a-hex-literal-made-llm-code-reviewers-raise-false-alarms/6ae24dc650f6fbe56d4c46ef/comment/6ae2d58627cf9dfd81c907eb#reply-6ae35300289414ae0e800634)

Useful design. Folding the literal first and giving the reviewer the folded value next to the original code would separate arithmetic from judgment. In my replies the models already got 3072 right and then judged it wrongly, so I would expect the folded variant to help less than a prompt that only says 3072 bits is fine. I have not tested that. On flag rates, I report balanced accuracy and pair-cluster intervals, and I have not run a flag-rate-matched comparison. I will note that as a limit.

[AI Developer](https://hashnode.com/@lycore)

[Oct 29](https://hashnode.com/posts/a-hex-literal-made-llm-code-reviewers-raise-false-alarms/6ae24dc650f6fbe56d4c46ef/comment/6ae2d58627cf9dfd81c907eb#reply-6ae36211ab46857a53df754b)

That distinction matters: if the models decode 3072 correctly and then misjudge it, the failure is in the security-judgment step, not the arithmetic. It also suggests a third arm could help: give the reviewer the folded value plus an explicit reference (for example, 'RSA 3072 meets current minimum guidance') and check whether the 21 'too small' verdicts flip, which would test whether it's a missing-knowledge issue or a miscalibration one. Noting the flag-rate limit sounds like the right call.
