Cerberus
A three-head reasoning and evaluation harness built on structured disagreement: one head proposes, one head critiques adversarially, and one head arbitrates and verifies.
Single-path generation has no internal check. Cerberus makes independent critique and explicit arbitration part of the reasoning architecture, so conclusions arrive with converged evidence instead of unexamined confidence.
Cerberus retains its research identity and published record, but is not a current development or runtime track. Its architecture and evidence are preserved; it is not offered as a live operational dependency.
Problem Space
A single model pass can be confidently wrong, and correlated evaluators fail together. Reasoning systems need independent construction, adversarial review, and a distinct arbitration step that weighs the evidence.
System Direction
Cerberus structures reasoning across three heads — proposal / construction, critique / adversarial review, and arbitration / verification — with independence requirements between heads and evidence convergence as the acceptance criterion. Security review is one application of the harness, not its identity.
Public Capabilities
- 01Head 1 — proposal and construction
- 02Head 2 — critique and adversarial review
- 03Head 3 — arbitration and verification
- 04Evidence convergence scoring
- 05Evaluator independence requirements
- 06Applied review workflows, including security analysis
Cerberus is described here at the architecture level. Head prompts, arbitration mechanics, security-sensitive applications, and internal evaluation content are not disclosed.
What Is Not Disclosed
Private implementation details, security-sensitive internals, and unreleased runtime architecture are intentionally not disclosed.