Catch the unsupported assumption.
Check an AI extraction before a missing value enters your model.
Begin the lesson“Annual rent: £180,000.
Lease term: five years.”
A source gap needs checking. Never silently fill it with zero.
AI for property / document extraction and human checking.
A missing assumption is not zero.
Invented source excerpt: “Annual rent: £180,000. Lease term: five years. Rent payable monthly.” The excerpt says nothing about rent-free periods.
Not on the strength of this excerpt. It supports the annual rent and term, but it does not state the rent-free period. Record that field as “not stated in the supplied excerpt” and check the relevant lease material before using a number. An empty source field is not evidence of zero.
The method you will practise
- Require each extracted value to carry a supporting excerpt or a clear “not stated” flag.
- Keep directly stated facts separate from derived figures and unresolved fields.
- Review the relevant source material and resolve the gap with the responsible person before using the assumption. A citation still needs checking against its source.
AI for property / document extraction and human checking.
The summary says nothing about rent-free months. Is “0” a supported extraction?
Choose what you expect. The feedback will help you test your starting idea.
Review two possible extraction outputs
Compare the cases below. Notice what changes, then read the explanation of why.
A fictional extraction exercise, not a lease interpretation or a claim about a particular AI provider.
An AI reply now cites the same short excerpt beside “0 months”. What changes?
Give me a hint
Require each extracted value to carry a supporting excerpt or a clear “not stated” flag.
Now try it in your software.
For a permitted, non-confidential practice document, require each extracted field to include supporting words or “not stated”. Keep derived calculations separate from directly stated facts. This is a fictional extraction exercise, not lease interpretation.
You can prevent an unsupported AI answer from silently becoming a model assumption.
Need help with your own example?
The Companion is where you ask about the task on your screen. Include your software version, what you changed and the result you expected.
Read this module’s Companion exampleThis sample checks your reasoning here. It cannot inspect or certify your work in another application.