Evidence Gate sits between retrieval and answer generation. Every claim the model wants to make is checked against a retrieved source first. No source, no claim — the model refuses to answer rather than guess.

Dual retrieval (semantic + keyword) pulls candidate sources for a query. Each is ranked by legal authority — Act, Rules, Circulars, Case Law, in that order. The answer is written only from what survives this gate, and every citation is re-validated against the original source text before it ships.

Tax and audit work carries real liability. Evidence Gate is the reason Tax Companion targets 0% hallucination rather than tuning for fluency — an answer with no citation is treated as no answer at all.