• Environment Pinning: Exact environment setup is required. Use lockfiles (uv.lock, bun.lockb) and document hardware assumptions.
  • Data Provenance: Document exactly where data comes from and any cleaning/filtering applied. If you make deliberate tradeoffs (e.g., dropping orphaned records), log them explicitly so successors don't mistake them for bugs.
  • Model Pinning: Specify exact checkpoints and API model versions (e.g., not just "Gemma", but the exact snapshot).
  • Nondeterminism: Explicitly document sources of nondeterminism (e.g., LLM temperature, sampling seeds).