By Laini Byfield

Practice

How to run small data systems day-to-day without turning people into collateral damage.

Data lineageMergingCommsAppeals

Operational guardrails by pipeline stage

Intake

  • Document source, fields, and intended purpose before loading
  • Validate schema; detect drift early rather than at payout time
  • Tag each load with a load date and source version

Matching and merging

  • Prefer stable identifiers; define fallback rules in writing
  • Flag collisions, duplicates, and ambiguous matches for review
  • Quantify match uncertainty; set review thresholds before running

Scoring and eligibility

  • Make rules explainable in plain language before they go live
  • Log rule version and cutoff assumptions with every run
  • Define what can be appealed and what evidence counts

Communication

  • Avoid false certainty — name what may update later
  • Provide a "how to correct" path, not just outcome notices
  • Use safe reporting; small-n suppression where re-identification is plausible

A practical checklist

  • Is the data necessary for this purpose — what can be removed?
  • Can a person contest the outcome — what is the timeline?
  • Can you trace an outcome to its source file and rule version?
  • What are the known error modes — what is the repair plan?

These four questions are a condensed version of the ETHICMAP cycle applied to a single run. If you cannot answer all four, the system is not ready to produce consequential outputs. See the full ETHICMAP cycle →