Validation

Validation is the evidence layer for generated model behavior.

Current validation support is deliberately narrow:

scenarios identify expected output cells and comparison rules;
generated values and oracle values are supplied as JSON mappings;
reports compare scalar outputs and preserve mismatch diagnostics;
optional oracle execution is available through the Python API for supported formulas workbooks.

Example report assembly:

modelwright validation report \
  --scenario tests/fixtures/synthetic_model/baseline_scenario.json \
  --generated-values tmp/generated-values.json \
  --oracle-values tmp/oracle-values.json \
  > tmp/validation-report.json

Oracle Limits

The pure-Python formulas oracle is useful for controlled workbooks, but it is not a universal Excel replacement. Real workbook evaluation has already shown structured-reference syntax that blocks this oracle. Cached workbook values can be useful evidence, but they are not the same as independent recalculation.