Current AI decision-support systems produce answers from a single model, offering no mechanism for deliberation, structured dissent, or cryptographic auditability. Regulatory frameworks such as NIS2 and DORA increasingly require organizations to demonstrate that AI-assisted decisions are traceable, auditable, and tamper-evident. We present ENLIL, a production system that convenes a council of nine specialized AI models to deliberate on each query in parallel, producing a consensus Decree signed with ML-DSA-87 (FIPS 204, NIST post-quantum standard). A reinforcement learning router dynamically allocates models based on domain classification and historical prediction error. Explicit dissent tracking surfaces minority positions in the final output, enabling senior reviewers to identify contested reasoning at a glance. The full nine-model deliberation completes in under 150 seconds; the resulting PDF is cryptographically signed and audit-ready. A tiered budget system (2, 4, or 9 models) makes the architecture economically viable from startup to enterprise deployments at approximately €0.20 per full-tier Decree. 237 automated tests pass at 100%. ENLIL is deployed in production at enlil-council.com.
Miguel Angel Concha Estrada (Mon,) studied this question.