The approach
Put the answer
to work.
What would you need to check before using it?
can still rest on
a contestable assumption.
I begin with a specific task and a decision someone actually needs to make. That gives the work a purpose and makes the cost of being wrong concrete.
AI can help retrieve sources, build calculations, examine alternatives and find what needs attention. Each contribution needs a check appropriate to the job. A compiler can test code. It cannot settle an economic assumption.
The useful result is something another person can inspect: what was used, what was done, what changed and what remains unknown.
A practical sequence
Show the work that matters.
Define the task.
Name the decision, the inputs and the standard an answer must meet. Compare workflows doing the same job under the same acceptance conditions.
Verify the result.
Read the source. Check the period and units. Execute the calculation. Use observable tests where they exist, and keep unsupported claims visible.
Change an assumption.
Vary something that could matter to the decision. Show which cases enter or leave, how the result changes and what you can no longer conclude.
Follow the consequence.
Ask what changes in the real system: useful work, cost, reliability, adoption or cash. Economic and investment judgments need evidence at those links too.
Try it on your next confident number
Make it
inspectable.
Reproduce the result. Change one economically relevant assumption. Show every case included and excluded, and explain what you can no longer conclude.
This is a way to expose a claim’s dependencies. It does not guarantee a correct answer. Some questions need a further observation, a different test or more data.
See the worked example