Use one experiment identifier across the run record, code reference, dataset version, artifact path, metric summary, cost entry, and decision note. Keep content versions immutable when possible and record every transformation that creates a derived result.
Create a lineage chain that survives folder changes
The experiment row should link a repository commit or release, an immutable input version, the executed configuration, and a run identifier from the tracking system. Raw outputs inherit that run identifier. Derived charts and summaries should state which raw artifacts and transformation code produced them. Stable object identifiers or content hashes are stronger than a copied local path that can move without notice.
Store provenance for imported fields. If a script extracts a metric from a log, record the script version, source artifact, extraction time, and resulting value. If a person corrects metadata, preserve the former value and reason. The ledger can then distinguish a changed experiment from a repaired description of the same experiment.
Verify retrieval in both directions
Sample important decisions and ask for every experiment they cite, then follow each experiment to the exact code, input, raw output, and measured result. Reverse the test by sampling artifacts and locating the hypothesis, owner, status, and consuming decision. Broken links, duplicate identifiers, and orphaned results become acceptance failures rather than future archaeology.
Permissions and retention belong in the lineage record. A stable link that the next operator cannot open is not operational evidence. Record the owning system, access group, retention rule, and deletion event where applicable. The buyer remains responsible for deciding which records may be shared and how long regulated or sensitive data should remain.
Where the service stops
Reality Contact, LLC implements evidence organization and retrieval, but does not certify scientific validity, reproducibility, research integrity, regulatory compliance, safety, authorship, or the correctness of any experimental conclusion. The buyer approves the record schema and storage boundary, verifies scientific interpretations, controls access and retention, and decides which evidence supports the next experiment, publication, or funding choice. This is technical record organization and implementation; it does not replace legal, medical, scientific, safety, compliance, or professional advice. A complete ledger improves traceability but does not certify that a result is valid, reproducible, safe, ethical, or suitable for publication or regulation.
Sources: MLflow Tracking documentation; DVC data and experiment management documentation.