Filters and the row pipeline

A mask describes where to discover columns, which rows to consider, how to validate values and what to return. Discovery and filtering are separate: header.anchors locates a schema while where evaluates records in that schema. With no anchors, where column names guide discovery. With neither, structural table discovery applies.

Processing order

1. Discover headers or bind an explicit source.
2. Classify the original row and apply row_policy.
3. Preserve original values for a possible rejection.
4. Replace null/empty values using configured defaults.
5. Apply explicit type rules and validate required fields.
6. Evaluate where against the effective values.
7. Route matches/rejects and project matched values.

Defaults cannot invent columns. Missing schema bindings are schema errors; invalid values in an existing column are row rejections. A default never replaces zero, false or a nonempty invalid value.

Predicates and projection

all, any and not compose comparisons. Omitting where selects every valid data row. select controls output independently of validation and filtering. Without select, a predicate returns its referenced columns; without either, all bound columns are returned.

Filter models and operators is the normative reference for field names, aliases, defaults, operator semantics and projection order. Validation demonstrates the complete pipeline.