How LexCore works: the review pipeline
Upload → parse → multi-agent debate → CitationGate → supervision → output. A walk through what happens between dropping a file and reading a verified finding.
Every review follows the same disciplined pipeline. Understanding it explains why LexCore's output is defensible in a way a raw LLM answer is not.
- 1Upload — you drop a contract (PDF, DOCX, RTF, or a scanned image) into a matter. Bytes are stored in encrypted object storage.
- 2Parse — the document engine extracts text into offset-anchored chunks with page mapping and heading detection. Scanned or photographed pages route through OCR (kept local for sovereign tenants).
- 3Debate — a bench of specialist agents (risk, enforceability, materiality, data protection, precedent, and more) analyses the document in parallel rounds. They argue: findings are produced and challenged, not asserted once.
- 4CitationGate — before any finding can surface, its supporting quote must be located verbatim in your document. No quote, no finding. A quote that doesn't match is dropped.
- 5Governance — findings about the same passage are aggregated by a reputation-weighted, commit-reveal step that resists sycophancy; genuine disagreement is recorded and escalated, never silently dropped.
- 6Supervision — the Review Level (LAM) sets how high a finding's confidence must be to auto-pass; the rest escalate to a Human Gate. A triage step ranks what survives by risk and materiality.
- 7Output — verified findings render with citation badges that jump to the exact page, a clause traffic-light heatmap, and optional visual risk maps.
Note
Where LexCore ends and the model provider begins is an explicit boundary — the point a security reviewer will ask about. In the cloud tier, agents call a frontier model to reason; in the sovereign tier, they call a local model, so no prompt or document text is sent to a third-party model provider. On a self-hosted deployment that extends to storage as well. See Deployment modes.
Related technical documentation
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.
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.