Methodology.
How MetaModel turns your inputs into calculations, assumptions, simulated reactions, and comparisons, and how it keeps them apart.
The short version.
Your facts go in. Tested code does the arithmetic. The AI model reads the inputs, explains the results, and challenges weak assumptions. Every figure shows where it came from, and every result can be repeated.
What do you enter?
You type in the facts about your campaign. You can also upload files. MetaModel suggests values, and you accept or reject each one. Nothing commercial counts as fact until you confirm it.
Each confirmed set of inputs is saved as a numbered version. Every output names the input version it used, so you can always see what a result was built on.
Who does the calculations?
The numbers come from code, not from the AI model. Money is stored in whole minor units, such as cents, and added up with exact decimal arithmetic. That code is tested.
Before an output is published, MetaModel checks that the totals add up and that each figure points to its source. It also records which version of the calculations ran. Run the same inputs through the same version and you get the same result.
The language model reads your inputs, explains the results, and challenges weak assumptions. It does not produce the financial figures.
How sure is a scenario?
Every scenario carries a label that says how much evidence stands behind it. The four labels are assumption-only, benchmark-informed, historically calibrated, and causally supported.
A range always says what it is based on. MetaModel shows probabilities or confidence intervals only when a sound statistical model and an evaluation support them. It makes no claim about forecast accuracy.
Where do the figures come from?
Each figure is labelled as customer-provided, sourced, or a modelled assumption. The source is recorded next to the figure it supports, and MetaModel checks that the source fits your market.
When MetaModel searches public sources, the search uses category and market terms only. It never sends the confidential text of your brief.
Are the audience reactions research?
No. Audience exploration is kept apart from the financial calculations. Its reactions are hypotheses for you to consider, not research findings.
The exploration can suggest an assumption, for example a conversion rate. That suggestion cannot become an input until you review and confirm it.
How does the AI part work?
The diagnostic follows a fixed sequence. It assembles your confirmed inputs and checks the structure and the arithmetic. It applies the marketing and commercial criteria and attaches references. Then it validates the output and publishes a versioned result.
OpenAI is the main AI provider and Anthropic is the fallback. The fallback gets the same confirmed inputs and must pass the same checks. MetaModel uses it only after a provider failure, a timeout, or repeated validation failures. It never uses it to work around missing evidence.
Written by Ravi R. Singh, Skill Pilot Advisory.