Separate AI interpretation from reproducible calculations
Separate AI interpretation from numerical calculation by keeping a verified input table, explicit formulas, reproducible outputs, and a narrative that refers to those outputs. A model can help propose the method, but its prose should not be the only place where the arithmetic exists. This separation makes errors easier to locate and correct.
Python’s decimal documentation explains decimal arithmetic and precision behavior. It is useful when exact decimal representation and deliberate rounding matter. Choosing a numeric library does not validate the financial meaning of inputs; unit, period, and denominator checks remain separate tasks.
Keep four layers independent
The source layer stores values as disclosed, with units and locations. The normalized-input layer converts scales under stated rules. The calculation layer contains formulas and rounding policy. The interpretation layer explains the result and its limitations. A change in one layer should not silently rewrite the others.
For a small research memo, a spreadsheet or short script is enough. The requirement is inspectability, not a particular tool. Someone else should be able to reproduce the output from the retained inputs without asking the model to generate the same answer again.
| Layer | Record |
|---|---|
| Source | Original value, unit, period, and locator |
| Normalization | Scale conversion and any exclusions |
| Calculation | Formula, precision, and intermediate values |
| Interpretation | Statement supported and assumptions required |
| Review | Independent check and changed dependencies |
Do not allow the interpretation step to invent missing inputs merely because the formula requires them. An unavailable denominator produces an unresolved calculation, not permission to estimate it without disclosure.
Hypothetical growth and margin check
Assume a fictional business reports revenue of 80 and 100 in two comparable periods. The growth formula is (100 - 80) / 80, producing 25%. If AI divides by the closing value instead, the result is 20%, which answers a different question. Retaining the formula exposes the denominator error.
Suppose operating profit rises from 8 to 12 in the same hypothetical periods. Margins move from 10% to 12%. That is an increase of two percentage points, or a 20% relative increase in the margin ratio. The narrative should specify which measurement it means rather than interchange the two expressions.
These values are hypothetical and do not represent an issuer’s results. The example shows that arithmetic correctness and language correctness are distinct. A correct calculation can still be described with the wrong unit or interpretation.
Check dimensions before precision
Write units beside every input and intermediate output. Dividing millions by thousands without normalization can produce a large error that no rounding policy fixes. Similarly, a quarterly numerator and annual denominator may require a different interpretation from a matched-period ratio.
Independently recompute consequential outputs using a different presentation, such as a manual calculation for a simple ratio or a separately checked spreadsheet formula. The goal is to challenge assumptions and references, not merely repeat the same code. Copying an erroneous formula into a second tool is not independent validation.
Then define rounding at the display boundary where possible, preserving sufficient internal precision for the task. Explain whether a difference is numerical rounding or a source-data revision. Avoid printing many decimals when the original disclosure is coarse; apparent precision should not exceed the evidence.
Connect corrections to the memo
Assign an identifier to each output and link dependent claims to it. If an input changes, rerun the calculation and reopen every sentence using that result. This prevents a corrected workbook from coexisting with an obsolete executive summary.
The method is unsuitable for producing an exact result when source values are incomparable or missing. In those cases, show scenarios with stated assumptions or withhold the calculation. Common mistakes include trusting a fluent arithmetic explanation, confusing percentage points with percentages, normalizing only one side of a ratio, and checking the numeric result without checking the sentence it supports.
FAQ
Should AI never be used for calculations?
It can assist with formula design and review. Material results should remain reproducible from verified inputs and explicit calculations rather than depend on an unrecorded verbal answer.
Does using decimal arithmetic make a model correct?
No. It addresses numerical representation. Source quality, compatibility, formula choice, and interpretation require their own checks.