CitationGate: grounding every finding
The trust core. A finding with no verifiable quote is rejected; a quote that doesn't match the source is dropped. This is why LexCore's output is traceable.
CitationGate is the mechanism that makes LexCore verification-first. When an agent proposes a finding, it must attach the exact clause that proves it. That quote is then located verbatim in your document before the finding is allowed to surface.
The rules
- A finding with no citation is rejected outright.
- Each quote is matched against the source — a whitespace-flexible exact match first, then a fuzzy word-shingle fallback for long spans on first verification. A match returns real character offsets and a match type (
exactorfuzzy). - Quotes shorter than 10 characters are rejected; a finding passes only if at least one citation verifies.
- Re-verification on edit is strict (exact only). If you change an amount or a term, the finding that cited the old text is correctly flagged stale rather than quietly re-matched.
- Every decision — verified, failed, passed, rejected — writes an audit entry.
Page-level anchoring
Parsing captures the page (and passage) for each chunk, so a citation badge jumps to the exact page in the source PDF — satisfying passage-level citation granularity, the high bar procurement checklists look for.
The one honest exception: absence
Some of the most valuable findings are about what a contract doesn't say — a missing indemnity, an absent liability cap. These structurally have no quote to gate. Absence findings carry a non-quote evidence type ("the playbook requires this and extraction found no matching provision"), bypass the quote gate by design, and default to high supervision because a false "missing" claim is costly.
Related technical documentation
Multi-agent debate (DyTopo)
A tiered bench of specialist agents produces and challenges findings in parallel rounds, ordered by a learning loop — so a claim survives scrutiny before you see it.
Review Level (LAM): sufficient supervision, made auditable
How much human review is enough for a task? LexCore turns that judgment into an explicit, auditable dial that drives how many findings escalate to you.
Clause extraction & the 41-category pack
Pull structured clause facts — type, verbatim text, and typed attributes — with the same grounding discipline as review.