The deliberative architecture
Quvant does not use a single AI model. It uses a process: three models with distinct roles carry out deliberation. GLM analyzes, Kimi contests and Gemini synthesizes. Claude validates the result blind, receiving only the original input and the final verdict: it performs a terminal check and determines one of the canonical outcomes (VALIDATED, DIVERGENCE or HALT). A reliability engine penalizes contradictions, not just consensus, and a vault preserves every piece of evidence in an auditable form.
3+1 Deliberation
3+1: three deliberators produce the analysis; a blind Validator verifies the verdict without access to the debate. The Validator receives only the original input and the Synthesizer's verdict, and is chosen from a different model family than the one that produced the analysis.
The process roles
Analyst
Regulatory analysis and coverage.
Critic
Searches for exceptions, gaps, edge cases.
Synthesizer
Integration and prioritization.
Validator
Terminal blind validation on Professional and Enterprise plans. Claude verifies the verdict blind, without access to the debate.
[Real Evidence Pack demo, coming soon]
Models are updated automatically. We do not publish specific versions: every analysis uses the best available at the time of execution.
The Validator is chosen from a different architectural family than the three deliberators: a validator from the same family would share the same blind spots and would not constitute a blind check. No provider shares two roles, so a systematic error from a single vendor does not propagate through the entire deliberative process.