Home / Five gates before an answer
Verification-first means the pipeline can refuse. Five gates stand between a legal question and a SallyIP answer — and each gate’s output is recorded.
Relevance thresholds decide when retrieval has failed: below them, the pipeline returns zero results rather than the least-bad passages. That zero propagates — no quotes to verify, no labels to guard, no entailment to grade — straight to a QUALIFIED or RESEARCH REQUIRED answer. Most legal-AI failures begin with weak retrieval nobody admitted; the threshold exists so the admission happens by construction.
The gates prove grounding, not truth. A VERIFIED answer means every proposition traces to real, supporting passages — the legal conclusion can still be wrong if the retrieved law is incomplete, superseded, or from the wrong jurisdiction. That is why entailment is measured separately (66.7% on the P0 run), why practitioner grading exists as its own section, and why the operating rule stays "no evidence, no assertion" rather than "verified, therefore correct".
What this proves: answers are grounded in real sources. What it does not prove: legal correctness — see why integrity ≠ correctness and the pending practitioner review.
It retrieves evidence first, verifies each quotation word-for-word, strips citations that point nowhere, grades whether each source supports its claim, and labels the answer VERIFIED, QUALIFIED or RESEARCH REQUIRED — publishing its miss rates either way.
Whether a cited passage actually supports the proposition attached to it. A real citation with no entailment is the most dangerous legal-AI failure, and the reason SallyIP measures entailment separately from citation existence.
An architecture where retrieval and verification gate every answer, with refusal as a valid output and published miss rates — operating rule: no evidence, no assertion.