run_log_schema_driftmap.md

Run Log Schema (Driftmap Public)

Purpose: standardize what to record every time a suite is run so comparisons are reproducible.

A run log is not the model output itself. It is the metadata and file references that allow a run to be reproduced and compared.

Required fields (minimum)

• run_id: unique run identifier
• run_utc: timestamp in UTC
• model: model name + version (as available)
• runtime: where it ran (LM Studio, Ollama, OpenAI API, etc.)
• suite_id: which suite was used
• suite_file: path or filename
• settings: temperature, top_p, max_tokens (if applicable)
• scorer: which rubric/scoring method was used
• output_file: results CSV filename
• notes_file: run notes filename (if used)
• receipts: hashes for suite file and output file (SHA-256 recommended)

Suggested structure (example)

• Suite inputs live in prompts/
• Outputs live in results/
• Notes live in sample_results/ (or a dedicated run_notes/ folder)
• Receipts can be stored as:
  • a .sha256 file next to the output CSV, or
  • a single append-only receipt log

Why receipts matter

If an output CSV is edited after the run, the hash changes. Hashes are the integrity layer that makes drift comparisons trustworthy.