RDR-009 Provisional Security & Supply Chain

Mandate Safetensors Serialization for Evaluated Machine Learning Checkpoints

Require non-executable tensor serialization formats (Safetensors or GGUF) for any model artifact achieving Verified status.

Recorded: 2026-10-22
Review Cycle: Quarterly (Next: 2027-01)

Context & Problem Statement

Legacy Python pickle-based model checkpoints can execute arbitrary code upon deserialization, posing an unacceptable supply-chain risk for automated evaluation pipelines.

Decision

Model artifacts hosted on public registries such as Hugging Face Hub must provide weights in Safetensors (or auditable GGUF) format to qualify for a ResourceRank score above 75 (Verified).

Consequences

  • Positive: Eliminates arbitrary code execution during weight loading and enables zero-copy memory mapping.
  • Trade-off: Older legacy checkpoints lacking Safetensors conversions are capped at Provisional status until converted and checksum-verified.