SafeGGUF 0.1.0

SafeGGUF 0.1.0

SafeGGUF validates GGUF model files before an inference runtime loads them.
This release provides a CLI, C API, Python bindings and container integration.

Features

• Structural and tensor-layout validation with checked arithmetic.
• llama-cpp and gguf-spec validation profiles, with bounded allocation,
  logical-work and scanned-byte budgets.
• Single-process admit publishing to content-addressed storage, with
  admission schema v2 recording effective policy and build provenance.
• Stable exit codes, JSON diagnostics, canonical error codes and optional
  metrics and structured logs.
• C APIs for paths and open file descriptors; Python bindings that bundle the
  native library when packaged as a wheel.
• Nonroot containers for Linux amd64/arm64, with signed release images and
  a digest-pinned Kubernetes deployment example.

Defaults and compatibility

The CLI and Python bindings default to llama-cpp and automatic byte-order
detection. C callers passing NULL options use the same defaults; explicitly
initialize options when using a caller-owned structure.

The llama-cpp profile rejects embedded NUL tensor names. Admission statements
use schema v2; inspection and error diagnostics use schema v1. API and schema
versions are independent of the package version.

Verification

Release assets include binaries for Linux, macOS and Windows, SHA-256 checksums,
a keyless signature bundle and an SPDX SBOM. See SECURITY.md
for signature and provenance verification.

A PASS verdict covers the selected structural and resource policy. Model
authenticity, weight behavior and inference compatibility require separate
verification. See the deployment guide and
admission contract.