Validate table types without coercing away evidence problems — what should the operator record?
A permissive parser can turn malformed numbers, ambiguous dates or unexpected arrays into plausible cells. That hides operational problems and can distort downstream joins. Use field-level type rules and preserve the original value where permitted so a reviewer can distinguish a transformation error from a limitation in the returned evidence.
Reject unsafe conversions and retain raw values when normalizing returned research fields.
Write field, accepted_type, nullable and conversion_rule for each working column, keeping identifier fields opaque even when they look numeric. A numeric metric contract should identify unit and accepted representation, while an identifier remains a string even if all inspected examples contain only digits.
Publish the type-contract revision and unresolved exception references; downstream operators should know which derived fields are safe to compare. A downstream calculation should consume validated typed fields and reject unresolved values when necessary, leaving raw evidence accessible only through its permitted reference path.Define accepted types
Write field, accepted_type, nullable and conversion_rule for each working column, keeping identifier fields opaque even when they look numeric. A numeric metric contract should identify unit and accepted representation, while an identifier remains a string even if all inspected examples contain only digits.
Create derived typed columns
Store parsed_value beside raw_value with parse_state, rather than replacing malformed input with zero, false or an arbitrary date. Keep parse_state invalid-format separate from source_state unavailable; the former describes your conversion and does not establish that the source omitted the value.
Test conversion boundaries
Include fixture values for empty string, null, unexpected object and formatted numeric text, checking that unsupported forms produce explicit exceptions. An unexpected array should fail a scalar-field conversion explicitly, not become a comma-joined string whose appearance conceals a structural mismatch in the response.
Review error distribution
Inspect actual failed rows by field and reason without inferring that parse failures mean the underlying source value was absent or false. Review errors by actual required fields and artifact revision so a display-only issue does not automatically invalidate unrelated identifiers or supported source links.
Release validated output
Publish the type-contract revision and unresolved exception references; downstream operators should know which derived fields are safe to compare. A downstream calculation should consume validated typed fields and reject unresolved values when necessary, leaving raw evidence accessible only through its permitted reference path.
Keep the boundaries clear.
- Reject unsafe conversions and retain raw values when normalizing returned research fields.
- Operational record: Define accepted types. Context badges do not certify access.
Start with a read-only check: ask your agent to report actual source coverage before collecting records or exporting an audience.
Primary references
Documentation reviewed 2026-11-15. Product capabilities, client availability and plans can change. Third-party marks identify their products, not an affiliation.