Computed fields

Language models read numbers well and calculate them badly. A compute field is calculated by code from other fields, so totals, ratios and installments are exact:

"net_salary":   {"type": "number", "integer": true, "max_digits": 10},
"installment":  {"type": "number", "integer": true, "max_digits": 10},
"dsr_percent":  {"type": "compute", "expr": "installment / net_salary * 100", "round": 2}

The pattern: the model extracts, code computes, the model explains.

Options

Option	Meaning
expr	The formula: field names, numbers, + - * /, parentheses, abs, min, max, round.
round	0 gives an integer, 1–10 rounds to that many decimals, omitted keeps full precision.
on, on_reject	Flow rules, as on any field.

• Fields named in expr become dependencies automatically.
• A computed field can use other computed fields (new_installment → total_installment → dsr_percent).
• Later fields see the computed value like any other value.
• A name that is not a field of the schema is rejected with HTTP 400.

Missing inputs

If an input is null or not a number, or the formula divides by zero, the result is null. The request does not stop by itself, because a missing value can be legitimate. Decide with a rule:

"dsr_percent": {"type": "compute", "expr": "installment / net_salary * 100", "round": 2,
               "on": [{"when": {"eq": null}, "do": "reject"}]}

Use review instead of reject when the value matters but should not stop the request.

Policy gates

A computed field in a later stage is a good place for checks that stop the request, because a reject only cancels fields of its own stage: the values extracted earlier stay in result for your records.

"age_gate": {"type": "compute", "expr": "age", "round": 0,
                 "on": [{"when": {"lt": 21}, "do": "reject"},
                        {"when": {"gt": 55}, "do": "reject"}]}

In the response

• The value is in result.
• Each calculation is in trace: {"do": "compute", "field": "dsr_percent", "expr": "installment / net_salary * 100", "value": 40.29}.
• Computed fields have no confidence: they are not model output. Protect their inputs instead (extract, votes, confidence rules).

See the credit assessment example for seven computed fields in one schema.