The Model Council: cross-checking AI to cut unsupported claims

When you ask a generic AI chatbot a question about a stock, you get one answer from one model. If that model hallucinates a profit margin, you have no way of knowing unless you manually cross-check the SEC filings yourself.

In investing, a hallucinated number isn't a "quirk of the technology"—it is a massive financial risk.

That is why Taufolio does not rely on a single AI model. We built the Model Council—a multi-stage, multi-agent pipeline designed specifically to reduce unsupported claims and improve accuracy.

The Architecture: A Multi-Agent Pipeline#

To build a professional-grade research platform, we had to move beyond the "one prompt, one answer" paradigm. Our Full Company Reports use a sophisticated pipeline powered by multiple APIs, external news verification, and several specialized AI agents working together.

While we keep our exact proprietary prompts confidential, here is a high-level look at how the pipeline operates:

1. The Data Gathering Agents#

Before any analysis begins, specialized data agents collect the primary sources. They pull the latest 10-K and 10-Q filings from the SEC's EDGAR database, retrieve earnings call transcripts, and integrate with multiple external APIs to fetch high-quality, verified market news (like Bloomberg or Reuters). This ensures the AI is constrained to a grounded reality, rather than relying on its pre-trained memory.

2. The Specialized Analysts (The Council)#

Once the data is assembled, it isn't fed into just one model. We use a multi-model approach. Three separate, independent "Analyst Agents"—often powered by different frontier foundation models from top providers—are given the exact same set of documents and the exact same research task (e.g., "Analyze the company's competitive moat").

Because these models have different architectures and training weights, they notice different things.

3. The Coordinator Agent#

This is where the magic happens. The three independent analyses are sent to a final Coordinator Agent. The Coordinator acts like a strict Managing Director at an investment bank. Its job is to compare the answers.

• Agreement: If Agent A and Agent B agree that margins dropped due to supply chain issues, and both cite the same paragraph in the 10-K, the finding is verified and included in your report.
• Disagreement: If Agent A claims margins dropped by 5%, but Agent B says 2%, the Coordinator will flag the discrepancy, re-check the source text, and if it cannot be definitively proven, the claim is rejected.

Why we use multiple news and data APIs#

We don't want the AI trying to remember what happened to a company last month. Memory creates hallucinations. By plugging the Model Council directly into live financial data APIs and premium news feeds, we force the AI to read current text rather than guessing.

The result? You get a report that is grounded in facts, cross-checked by multiple "analysts," and verifiable via inline citations.