Generation Contract Corrections

English · Русский

These corrections are included in stable 1.6.0. Regenerate affected fixtures and review differences when upgrading. Package version and DatasetSpec schema version are separate; the current specification schema is 1.1.

Changed Results

  • Formula dependencies are computed before their consumers, regardless of list order. Cycles and multiple assignments to the same formula target fail early.
  • Constraints and relationships marked rejected are ignored by generation and validation. Existing inferred rules in an explicitly supplied specification remain executable; the agent workflow still requires fingerprint approval. Remove or revise active aggregate mappings that depend on a rejected link; they fail validation instead of being silently skipped.
  • Nullable foreign keys preserve generated nulls. One-to-one capacity counts only non-null child references. Required identifiers cannot be empty or null.
  • String-pattern lengths describe the entire string, including any synthetic prefix. Short non-sensitive strings use random characters where a prefix does not fit. A required string cannot have a zero maximum length.

Regenerate affected fixtures and review resulting differences. A fixed seed remains reproducible within the recorded generator version and environment; it does not promise identical bytes across this upgrade.

Publication And Validation

Valid, edge, and load-test generation must pass all core schema, relationship, constraint, and privacy checks before returning rows or publishing output. Turning off report sections does not disable these generation checks. Business rule mutations are checked again before publication. Failures preserve existing output and do not publish a new successful bundle.

Explicit mixed and negative generation still publish controlled-invalid data with validation evidence. Parquet uses explicitly marked scalar columns when intentionally malformed cells cannot fit native Arrow types. The bundle reader restores those cells for validation; consumers reading Parquet directly must inspect the apa.scalar_encoding field metadata. Ordinary valid-mode values incompatible with declared types still reject the whole export without partial publication. Row privacy remains mandatory in every mode. The RC5 CLI reports validation_failed with exit 1 for intentionally invalid results while retaining the representable output, manifest and report. Explicit mode/ratio overrides must be reflected in effective execution evidence; an invalid artifact is not a successful validation. Standalone validate now inspects row values when its privacy section is enabled, in addition to checking specification safety. CSV numeric values are interpreted using their declared numeric type before privacy classification. This decoding exception applies to revalidation only; generated formula and business-rule strings retain the strict pre-export privacy check.

These checks detect recognizable sensitive strings using the existing synthetic namespace policy. They do not certify anonymity or prove that data has never come from a source. Keep the source-row checks and organization-specific review.

The solver does not search for a solution to arbitrary interacting rules. If later temporal, conditional, or aggregate changes invalidate an earlier rule, the operation fails; revise the specification rather than consuming partial output. Numeric sampling remains uniform between configured bounds and foreign key assignment remains round-robin among non-null child rows. These are not claims of preserving correlations or source fan-out distributions.

Additive validation utility evidence

New reports add utility with the canonical effective-spec hash, active native relationship/temporal declaration counts and executed/not-declared/not-run status. Temporal metrics add generated checked/failing row counts. Existing validity, settings, section errors and relationship metrics retain their semantics. Old saved reports without utility remain accepted. JSON consumers that enforce exact keys must admit this new optional field.

Generated row values, DatasetSpec versions and existing profile/spec hashes are unchanged. Serialized validation reports and their manifest output digests do change. Compare replayed rows separately from report byte identity across this upgrade; the new contract fixtures record the intentional report changes. Unknown source fidelity, fan-out and missing relationship coverage remain unknown. No source queries, new authorization, dependency, package publication or completed release-candidate acceptance is implied.

Parquet distinctness evidence

Reprofile Parquet sources to obtain bounded non-null distinctness. Old profiles retain unknown values. Newly profiled scalar columns may have exact or lower_bound unique-ratio evidence using existing schema fields; changed profile fingerprints require fresh plan review. This adds no primary-key nomination, identifier pool or source key reuse. Missing null metadata still blocks automatic spec inference. Existing dataset/profile schema versions remain unchanged.

Long fractional datetime intervals

Datetime range sampling now computes its whole-second span with integer arithmetic. Very long fractional intervals previously could round the maximum admissible offset up and exceed the upper bound by a microsecond. Replaying such an affected specification may produce different values after upgrading; compare the manifest's generator version. The existing second-resolution policy, missing-bound fallbacks and schema version are unchanged. This does not add monthly/calendar inference or certify source-period fidelity.

A partial date/time bound must also be compatible with the existing missing endpoint defaults (minimum 2020-01-01, maximum 2025-01-01). A maximum before 2020 or minimum after 2025 now rejects instead of silently clamping an inverted interval. Supply both ordered bounds for earlier or later periods. Failure leaves previous dataset artifacts unchanged and publishes no new manifest.

CSV numeric replacement evidence

Reprofile CSV folders to replace cache format 7 with format 8. Ambiguous numeric phone-like fields may now include optional numeric_storage_type; sensitivity and masked patterns are unchanged. Old ambiguous profiles cannot opt into a numeric replacement without fresh storage evidence. Independently authored numeric_measure decimal ranges use the existing infer-spec --decimal-fields contract. Re-review changed profile/spec fingerprints. In an awaiting-approval workspace, copy the reviewed field while preserving planned generation settings, run agent-review and approve the new fingerprint. Do not mutate an already approved plan; replan changed source/profile evidence. DatasetSpec version 1.1 and unaffected legacy hashes remain unchanged; see profiling.

Query-summary batching

Query-source summary batching changes trusted internal SQL and request counts, not profile/spec schemas. Fixed aggregate evidence keeps the same profile fingerprint; there is no silent reclassification or saved-plan promotion. Existing statement, scan, result and time limits remain enforced independently. Real source changes still require normal profile review.

Optional conditional formulas

Formula rules add optional when with the existing typed Condition semantics. Absent/null predicates serialize as omitted and retain prior hashes and seeded behavior. Reports add eligible/evaluated/skipped counters only for conditional formulas; strict consumers should accept these additive counters. DatasetSpec stays at 1.1. Reviewed conditional branches change the rules/spec fingerprint and require renewed plan review. Conditional writer conflicts reject before publication; no existing source access or privacy permission is expanded.