Relational Synthesis Contract¶

Relational synthesis preserves reviewed, executable structure. It does not reconstruct source rows, identifiers, totals, or rare values.

Evidence-Bounded Synthesis¶

The system does not magically discover the complete domain model from a schema or a sample. It records bounded observations, keeps their provenance, and turns unverified conclusions into hypotheses that require review.

The conceptual pipeline is:

SourceBundle
  schema + constraints + examples
        ↓
EvidenceProfile
  safe observations + provenance
        ↓
SemanticHypotheses
  candidate relationships and rules
        ↓
human review
        ↓
ReviewedDatasetSpec
  executable generation contract

These artifacts have deliberately different authority:

Artifact	Meaning	Generation authority
SourceBundle	Input schema, constraints, examples, and sampling metadata	None
EvidenceProfile	Safe observations supported by bounded evidence	None
SemanticHypothesis	Deterministic or AI-proposed relationship/rule	None until reviewed
ReviewedDatasetSpec	Explicitly accepted and executable contract	Yes

Provenance And Confidence¶

Every non-declared relationship, distribution, or business rule should retain an evidence envelope containing, where applicable:

• origin: declared, observed, or inferred;
• status: candidate, reviewed, accepted, or rejected;
• confidence: low, medium, or high;
• sample size, sampling method, and time window;
• rows checked and observed violations;
• source/profile fingerprint and explicit assumptions.

Confidence describes how strongly the conclusion is supported by the available evidence. It is not a claim that the rule is universally true. For example, observing invoice.total == sum(lines.amount) in 200 rows with no violations supports a medium-confidence hypothesis; it does not prove the domain rule.

AI may rank or explain hypotheses, but it cannot increase their confidence, change their review status, or authorize generation. Missing evidence remains unknown and must either be reviewed as an explicit assumption or remain outside the executable contract.

Preservation Matrix¶

Property	Preserved contract	Not preserved	Evidence
Foreign-key graph	Reviewed parent/child edges, supported cardinality, null policy, and referential integrity	Source key values, row membership, or source ordering	DatasetSpec relationships and relationship validation
Distribution shape	Configured safe ranges, ratios, bounded categories, and order of magnitude	Exact histograms, rare categories, source totals, or a statistical privacy guarantee	Effective field distributions and validation report
Temporal dependencies	Reviewed ordering rules, bounded ranges, and supported lifecycle constraints	Original timestamps, event sequences, or per-source-row timing	Executable temporal rules and business-validation report
Business invariants	Supported formulas, conditions, partitions, coverage, foreign keys, and aggregate mappings	Free-form prose rules or unvalidated model conclusions	Effective rule fingerprint and deterministic validation results

Approval Boundary¶

Only reviewed DatasetSpec relationships and structured business rules affect generation. Deterministic mining and AI advisors may propose candidates, but a proposal has no generation authority. Low-confidence, contradictory, stale, or unsupported candidates remain review warnings or are rejected.

Scaling And Privacy¶

Synthetic row counts and measures may scale while retaining approved relative shape and order of magnitude. Scaling must use synthetic values and explicit rules; it must not copy source totals or preserve sensitive outliers. Exact-row reuse checks are a safety backstop, not proof of anonymity.

Numeric distributions accept a bounded scale_factor from 0.1 through 10.0. Sensitive numeric distributions require an explicit non-identity factor; generation scales the approved bounds before drawing values and records the effective distribution in dataset_spec.yaml.

Publication Rule¶

A dataset may claim an invariant only when deterministic validation executes that invariant successfully. Failed or unsupported checks must block valid publication or appear as explicit controlled-invalid evidence. The manifest, effective spec, and validation reports form the review record.