Model input provenance

Model input provenance records where every material spreadsheet or code input came from and how it was prepared. For each input, retain its source exhibit, owner, date, period, currency, unit, sign convention, transformation, assumption state, and version. A cell address alone is not provenance because the upstream cell may have been typed from an unretained message.

Use one labeled input table for the formatted workbook and readable Python re-performance. Both views should consume the same values and produce the same control totals. If a $2.4 million balance enters Python as 2,400, record whether the model uses thousands and show the conversion. If a rate changes, preserve the earlier value and source in the revision trail.

Separate supplied facts from assumptions and management estimates. A model can recompute interest, allocation, or sensitivity under those inputs. It cannot establish that an estimate is reasonable, that recognition is correct, or that an owner approved the value.

Review provenance before reviewing formulas. A perfect formula applied to the wrong entity, period, or currency gives a precise wrong result. Stop when the two model views disagree, trace the first difference to its input or logic, and record the correction rather than choosing the preferred output.