Please check the build logs for more information.
See Builds for ideas on how to fix a failed build, or Metadata for how to configure docs.rs builds.
If you believe this is docs.rs' fault, open an issue.
harn-guard — management and inference for downloadable on-device models.
This crate is the management layer: a purpose-tagged catalog of upstream,
already-hosted models ([catalog]), an on-disk store that installs and
verifies them ([store]), and purpose-scoped selector resolution
([resolve_dir]). It hosts nothing, bundles no weights, and makes no network
calls itself — the CLI downloads from the catalog's upstream URLs on the
user's machine and hands the bytes to [GuardStore::install] for SHA-256
verification + atomic install.
The heavy inference runtime (ONNX) lives behind the off-by-default neural
feature, so the default binary never links a model runtime. When built
without it, [load_classifier] is a no-op returning None and the runtime
keeps using the always-available built-in heuristic classifier. When built
with it, a host installs [load_classifier] into the runtime's lazy loader
seam (harn_vm::security::set_injection_classifier_loader), which fires the
first time a local-ml policy scores untrusted content.