How LexCore works: the review pipeline
Upload, parse, multi-agent debate, CitationGate, supervision, output. What happens between dropping in a file and reading a verified finding.
Every review goes through the same pipeline. Once you know how it works, it is clear why a LexCore finding holds up where a raw LLM answer does not.
- Upload. You drop a contract (PDF, DOCX, RTF, or a scanned image) into a matter. The file is stored in encrypted object storage.
- Parse. The document engine extracts the text into offset-anchored chunks, with page mapping and heading detection. Scanned or photographed pages go through OCR, which stays local for sovereign tenants.
- Debate. A bench of specialist agents (risk, enforceability, materiality, data protection, precedent and others) analyses the document in parallel rounds. Findings are proposed and then challenged, not asserted once.
- CitationGate. Before a finding can surface, its supporting quote must be found verbatim in your document. No quote, no finding. A quote that doesn't match is dropped.
- Governance. Findings about the same passage are aggregated by a reputation-weighted, commit-reveal step that resists sycophancy. Real disagreement is recorded and escalated, never silently dropped.
- Supervision. The Review Level (LAM) sets how confident a finding must be to auto-pass. Everything else escalates to a Human Gate. A triage step then ranks what survives by risk and materiality.
- Output. Verified findings appear with citation badges that jump to the exact page, a clause traffic-light heatmap, and optional visual risk maps.
Note
There is an explicit boundary between LexCore and the model provider, and a security reviewer will ask where it sits. On the cloud tier, agents call a frontier model to reason. On the sovereign tier, they call a local model, so no prompt or document text goes to a third-party model provider. A self-hosted deployment extends that to storage as well. See Deployment modes.
Related documentation
CitationGate: grounding every finding
A finding with no verifiable quote is rejected, and a quote that doesn't match the source is dropped. This is what makes LexCore's output traceable.
Multi-agent debate (DyTopo)
A tiered bench of specialist agents proposes and challenges findings in parallel rounds, in an order set by a learning loop. A claim has to survive 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 setting, recorded for audit, that decides how many findings escalate to you.