ADR-0005: Foundry produces semantic values

Status

Accepted

Context

Each entity has many attributes with CDM data types and a purpose. Some are identity and audit fields that must line up with the generation clock and the key space; others are semantic fields such as names, descriptions, and amounts that should look plausible for the industry.

Decision

Split each record. The engine stamps system fields: primary keys, audit timestamps by purpose, state codes, and foreign keys drawn from the identity pool. Azure AI Foundry produces the semantic fields through a structured-output call whose JSON schema is derived from the entity's semantic attributes. The engine validates and coerces every returned value to its CDM type.

Consequences

• Keys and timestamps are correct and cheap, because the engine computes them rather than asking a model.
• The model focuses on the values it is good at, which lowers token cost and improves realism.
• Generation is repeatable in structure and distribution, not bit for bit, because model output is not deterministic.

Alternatives considered

• Ask the model for every field, including keys and timestamps. Slower, costlier, and less correct.
• Rule-based generation for all fields. Fast and offline but less realistic for free-text fields. The offline providers in the hosts use this approach for dry runs.