curation-discipline

The staging-gate discipline for AI-assisted data curation — agents collect into staging, a read-only validator checks, a human approves, and only then does anything reach the canonical store. Use when building or reviewing a pipeline where an LLM agent contributes data to a database, knowledge base, or curated dataset, when deciding which agent may write where, or when a curated dataset has developed duplicates, unattributed rows, or silent gaps.

How to record where curated data came from and how to detect that a source was already imported — provenance coordinates instead of text comparison, bidirectional reconciliation of what an agent extracted, and the deliberate-discard list that makes an agent's judgment reviewable. Use when the same source may be submitted twice, when an agent's output must be checked for omissions, when deciding what provenance fields a schema needs, or when a curated dataset has developed duplicates.

A catalogue of data-pipeline failures that produce no error message — joins that match nothing, dry runs that were real writes, fan-out that duplicates rows, guards that die on an encoding error, and tools that were written but never merged. Use when a pipeline reports success but the data is wrong, when reviewing an import or sync script, when deciding what to verify after a batch write, or when designing a check that has to catch something the runtime will not.