Before deciding whether two records describe the same entity:
- Check an AI extraction against the original document; keep the exact name and source reference.
- Preserve stable source/record IDs and original labels. Do not overwrite with normalized text.
- Record scorer, processor, version and shortlist limit. A score is not a calibrated probability.
- Inspect extra words, numbers, legal suffixes, locations and tied candidates.
- Check independent approved identifiers and original evidence. Names alone do not establish identity.
- Record reviewer, evidence, decision and reason, including unknown/no match.
- Preserve a reversible audit trail before any separately authorized system change.
- For production: validate schemas, missing/non-string/blank inputs and scale; evaluate candidate recall and decisions on labeled in-domain examples. This toy script does none of those production evaluations.
