Eligibility assistant

The result is computed, not written. The schemes retained, the amounts and the sources come from a deterministic engine applying published eligibility rules — same inputs, same outputs, no model and no network.

A model then does one job: write the explanation. It receives the already-computed result and chooses nothing. Its text is discarded if it contains a figure absent from the data — in which case the computed result stays, because it never depended on the model.

Amounts are absent from the indexed corpus. A model cannot restate a figure it never received: the error becomes impossible rather than improbable.

Income bracket is not asked, deliberately. The data carries labels — "modest", "intermediate" — with no numeric threshold, and "modest" does not mean the same income in Luxembourg and in Romania. Making you pick a category the data cannot define would discard schemes on a guess. Means-tested schemes are therefore all shown, each with its own condition, and you judge.

The detail field is the only free text on this page. It passes through a prompt shield before reaching the model: an attempt to hijack the assistant is discarded, and the rest of the request is handled normally. If the shield is unavailable, the detail is discarded too — a control that goes dark must not open the door it was guarding.