Deliberative governance and anti-sycophancy
Reputation-weighted, commit-reveal aggregation that stops the AI agreeing with itself, and records dissent instead of hiding it.
Multi-agent systems have a known failure mode: sycophancy, where agents converge on the first or loudest answer. LexCore's agents already commit their findings independently and blind. On top of that, we add a reveal-and-aggregate step.
- Reputation. An EigenTrust-style calculation weights agents by how much they corroborate each other over time, seeded by track record. How many findings an agent happens to file does not count.
- Commit-reveal aggregation. Findings about the same passage are clustered by position. The reputation-weighted majority becomes the consensus, and minority dissent is recorded, never dropped.
- Contest detection. When conflicting positions on one passage both have real support, the passage is flagged as contested and escalated to an arbiter. If that fails, it goes to a Human Gate.
Trust
The recorded dissent is the anti-sycophancy audit trail. If two agents disagreed about a clause, you can see that they did, and how it was resolved.
Related documentation
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.
Human Gates and the triage worklist
Keeping a lawyer in the loop without burying them: what escalates, and how the rest is ranked by risk and materiality.
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.