Lineage: proving where an answer came from

An agent is a fluent amplifier for whatever data it reads. If you cannot say which rows, from which run, under which contract, produced a given answer, you cannot defend that answer to a model-risk reviewer or fix it when it is wrong. So you write a data contract for every dataset the agent depends on, test it on every load across the six quality dimensions, quarantine the rows that fail, stop the index refresh when the batch fails, and record lineage at table and column level as machine-readable events. The catalogue is where owners, definitions, quality signals and lineage live together so that someone other than you can answer "where did this come from?" next year.

• Quality is measured per dimension, per check, with a threshold and a severity. "The data is good" is not a measurement.
• A contract is an agreement with a named owner, not a schema file. ODCS gives it a portable shape.
• Gate the consumer, not the whole estate. Block the agent index refresh; keep serving the last good snapshot.
• Lineage is evidence only if it is captured at run time and links to the exact snapshot the agent read.
• Every index document carries a source reference (dataset, primary key, row hash, extract time, run ID, contract version). That reference is what makes the trace possible.