Skip to content

Anti-patterns

  1. The dumb mirror. Wholesale source replication that adds nothing. Scale is fine — millions of Things are fine — when the repo adds identity, normalization, provenance, and an agent-native query surface. Large canonical semantic artifacts such as imagery and point clouds live in object storage with a hash and durable locator.
  2. Assuming either that every source identifier is canonical or that none can be. An ordinary record key does not become real-world identity by convenience. Conversely, demoting a mature, stewarded domain registry to mere evidence can create a private competing identity with worse interoperability and correction behavior. Evaluate the provider’s definition, identifier lifecycle, matching, stewardship, and user workflows; adopt it when it answers the competency questions, and depart only for a demonstrated benefit.
  3. Relationships in names. Permit/issued-by-ncdeq/... lies the day delegation changes.
  4. One confidence float to rule them all. Source quality, model score, calibration, and belief are different things; use method classes, evidence, and opinions on binomial propositions only.
  5. Epistemologizing the furniture. “Code 00610 means ammonia” is a mapping, not a debatable proposition. Reserve propositions for genuinely contestable binary statements — and discover them from repeated practice rather than inventing them in a conference room.
  6. Designing against hidden internals. Model with the documented platform surface; internals change under you.
  7. Premature splits. Extract shared-provenance repos or split identity from geography when operational pain demands it, not before.
  8. Skipping the boring artifacts. The naming contract, the field-level source mapping, the anti-inference list are the ontology; the Shapes are their compilation target.
  9. Fabricated data in public repos. Seed with a small real slice, or stay private until it’s real.
  10. Big-bang scope. The vertical slice exists because ontologies are only proven by answering real questions against real data with real provenance.
  11. Wrefs in string fields. No edge, no backlinks, silent staleness — ever.