What the models do
The models generate structured specialist outputs from the same evidence pack, operating within editorial domain boundaries that define what each Panel specialist considers in scope. HELIOS is not allowed to drift into PSYCHE's territory; ATHENA is bounded to fundamentals and valuation. This domain separation is enforced by protocol, not left to the model's own judgment.
The five SENIOR instances then each independently evaluate the fused Panel picture using different model architectures: Google Gemini, OpenAI GPT, Anthropic Claude, DeepSeek, and xAI Grok. These are not stylistic variants of the same underlying model -- they are architecturally distinct systems that may weight evidence differently, surface different considerations as primary, and reach different conclusions from identical inputs.
What the models cannot do
The models do not replace fund prospectuses, tax law, or your personal suitability assessment. They do not know your tax jurisdiction, your portfolio's existing exposures, your liquidity needs, or your investment time horizon. Every model output in the Cognitor pipeline is general information -- a structured interpretation of the evidence available at the time of production.
The models also do not "pick winners." There are no buy or sell signals in the Cognitor output. The framework is built around scenario assessment, lens-level agreement and disagreement, and evidence quality mapping -- inputs to your thinking, not conclusions for you to execute blindly.
Why architectural independence matters more than "smarter text"
If every step of your research pipeline shares the same underlying architecture, you can get fluent, confident, well-written wrong answers. The correlation problem is invisible -- you cannot see it in the output quality. A single highly capable model evaluating the same evidence from five different angles will still share the fundamental blind spots of that model architecture.
Cognitor's independence layer is the key safety mechanism: six Panel lenses run within domain-bounded protocols, then five SENIOR instances use genuinely distinct model architectures. This means that when four SENIOR verdicts agree and one disagrees, that disagreement is more likely to be informationally meaningful -- it is not just one model contradicting itself. When all five agree under architectural diversity, that consensus is structurally more robust than five copies of the same model unanimously agreeing.
How to evaluate any "AI research" product
Ask four questions. First: are disagreements visible in the output, or is everything collapsed into one clean recommendation? If all you see is a final answer, the disagreement was hidden from you. Second: is the protocol fixed and documented, or does the output character change based on unspecified factors? Third: is the ETF universe explicit and curated, or does the system claim to cover everything equally well?
Fourth, and most important: are the models architecturally distinct or are they rebranded copies of the same system? If the answer is copies, you have a summarizer with extra steps -- not a deliberation stack. Cognitor publishes all five SENIOR outputs in the dossier so you can see the divergence directly, read the minority verdict, and judge for yourself whether the consensus is earned or merely uniform.