How can generated records preserve monetary precision across export formats?

A generated release can specify a monetary schema before choosing its export formats. Fields can include `amount_decimal`, `currency_code`, `currency_scale`, `rounding_rule_id` and `conversion_rate_id`. Adjustment lines can link to a conversion record when source and reporting currencies differ.

For JSONL, decimal amounts can be encoded as strings with a documented grammar. For Parquet, the schema can use decimal logical types with declared precision and scale. The accompanying manifest can define how each representation maps to the same canonical amount.

Rounding rules should state whether rounding occurs per line, per tax component or after aggregation. A generated task can then contain a legitimate rounding remainder without forcing the model to invent a balancing explanation. Unsupported currencies or conversions can be excluded explicitly from the pilot scope.

Import validation can compare canonical totals before and after serialisation and reject undeclared scale changes. Tool responses should use the same conventions as the exported records. This lets a lab attribute a numerical failure to agent behaviour or data handling with clearer evidence.