# Dependency security review
Run `cargo audit` as an informational local gate. Vulnerability findings block
release work until resolved or explicitly reviewed. Maintenance-only warnings
may be accepted temporarily when the dependency is transitive, no supported
upgrade removes it, and the warning does not describe a known vulnerability.
## The numerical dependency
FerricML has two runtime dependencies: `faer` for the dense decompositions and
`sha2` for artifact digests. `faer` replaced `nalgebra 0.34.2` on 2026-07-27,
and the reason was **correctness, not performance or supply chain** — the
previous backend returned an SVD that does not reconstruct its own input on
exactly-rank-deficient tall designs. The changelog carries the measurement. That
matters here because it inverts an argument this document used to make.
Measured on the registered `aarch64-apple-darwin` runner, the shipping graph
(normal plus build dependencies, development dependencies excluded) is 56
crates. Every one of them can be taken under MIT, with a single build-time
exception: `unicode-ident` is `(MIT OR Apache-2.0) AND Unicode-3.0`, so the
Unicode licence applies alongside whichever of the pair is chosen. Nothing in
the graph is copyleft.
## Reviewed maintenance warning
- `RUSTSEC-2024-0436` applies to `paste 1.0.15`, a build-time macro crate
FerricML never names directly. The advisory reports that the crate is
unmaintained; it reports no vulnerability.
- **How it is reached, read off the dependency graph rather than assumed.**
Before the backend change it came in through `nalgebra 0.34.2` and
`simba 0.9.1`. It now comes in through `faer 0.24.4` and `gemm 0.19.0`, and
through the four `gemm-*` kernel crates beside it. The path changed; the
presence did not.
- **The earlier assessment is withdrawn in its premise and kept in its
conclusion.** This review used to reason that replacing the numerical backend
solely to drop this macro "would carry disproportionate numerical and
compatibility risk". That reasoning is now known to have been backwards: the
backend was replaced, and it was replaced because keeping it was the numerical
risk. The conclusion survives on a plainer fact than the one it used to rest
on — the new backend depends on `paste` as well, so no backend choice
available to FerricML removes it, and a swap made for this advisory alone
would buy nothing.
- **The general lesson, since this review got it wrong once.** "Changing this
dependency is riskier than living with the warning" is a claim about the
dependency's quality, and it needs evidence of that quality rather than the
inertia of an incumbent. Nothing had measured the incumbent when that sentence
was written.
- Review again by 2026-10-24, or earlier when the numerical dependency updates.