Read this before relying on candor for anything security-critical.

candor is an advisory static analysis, not a sandbox or an enforcement mechanism. It tells you what code appears to do; it does not stop code from doing anything. It also has documented blind spots — calls it cannot resolve are marked Unknown (dynamic dispatch over non-local traits, function pointers, closures through callbacks), its classifier only knows a curated set of crates, and macro-generated functions are skipped. See CRITIQUE.md and BACKLOG.md.

Do not use candor as a security control. A &Fs token is a lint convention, not a gate — you can still call std::fs without it. For real, compile-enforced capability isolation use cap-std; for supply-chain API restrictions use Cackle. candor complements these as a visibility layer; it does not replace them.

Where candor does help your security posture: surfacing the effect surface (what reaches the network / filesystem / spawns processes), and the regression guard (catching a change that makes a previously-pure function start doing I/O). Treat both as defense-in-depth signals, not guarantees.

The bug class that matters most here is a false negative: candor reporting a function as effect-free (or missing an effect) when it actually performs that effect — a soundness hole. Those are prioritized, because a checker that's silently wrong is worse than no checker (PRINCIPLES #1).

Please report via a GitHub issue with a minimal reproducer (the smallest function + crate that's misclassified). If you'd rather report privately, email tom@polymorphism.co.uk.

candor is a prototype pinned to a specific nightly toolchain. There is no formal support window; fixes land on main.