What the AI does, and what it is not allowed to do.

The role of the language model

A generative model is used for two things: reading the question, and composing an answer from material retrieved out of the connected community model. It is not the source of the facts in that answer, and it is not the authority on whether a statement is true. The connected source is.

• Values come from connected sources, not from the model's own knowledge.
• Answers are scoped to one community's model, so a citation is possible.
• Where retrieved material does not support a conclusion, the platform reports that instead of composing one.

How claims stay traceable

Substantive claims carry the document, dataset, layer or indicator they came from. Measured values, values derived by calculation and stated assumptions are labelled differently and stay distinguishable all the way into an exported report, so a reader can see which kind of number they are looking at.

Confidence and gaps

Confidence is reported with the reason behind it, usually the coverage and refresh frequency of the sources involved. Declared gaps are fields in the model rather than notes someone remembered to add, so every answer inherits the gaps of the data it rests on, and an exported report places them before the conclusions.

Known limits

• A language model can misread a document or attach a figure to the wrong area. The citation is shown next to the claim so this is checkable in seconds rather than discoverable months later.
• A model built from documents inherits their errors. TwinGov surfaces disagreement between sources rather than averaging it away, but it cannot know which source is right.
• Access and coverage analysis is only as good as the network and population data connected. Where those are partial, the analysis says so and does not extrapolate.
• Forward-looking projection is not part of the product. Alerts operate on observed series.

What TwinGov does not do

• It does not make decisions, approve investments or issue permits.
• It does not rank options as a recommendation the institution is expected to adopt.
• It does not profile individuals, and it is not designed to hold personal records.
• It does not fill a gap in the data with an estimate presented as a measurement.

Human responsibility

Every conclusion TwinGov produces is an input to a decision that a named person takes and remains accountable for. That is not a disclaimer added for safety. It is how the product is designed, and it is why the comparison output states a difference rather than a verdict.