Building DisconfirmAI: An AI investment research workspace that looks for counter-evidence first

When you use a general-purpose LLM to analyze a stock or read an earnings report, it tends to agree with whatever premise you put in the prompt. If you ask why a company looks undervalued, it will assemble a convincing bull case. In financial research, that habit is risky.

I built DisconfirmAI around a simple rule from Karl Popper and value investing: before building conviction in a thesis, look for the specific evidence that would prove it wrong.

Here is how the workspace is structured:

The Daily Reading feed curates AI industry and market developments from primary sources, keeping original links and publication dates. Editorial annotations use three fixed labels:

In the Research workflow, questions about earnings releases, multi-year financials, and SEC 13F institutional holdings (calculated strictly by share-count changes rather than market-value fluctuations) are split into three sections: Verified, Inferred, and Doubtful. It does not output buy/sell recommendations or price targets.

In Investment Framework, you write down your circle of competence, entry rules, and the exact conditions that would force you to revise or exit a position. Radar scans those conditions on a schedule. When nothing material changes, it stays quiet and logs scanned-and-empty in the audit trail.

Everything below is open in the browser without registration:

If you analyze tech or AI companies or build financial research workflows, I would love your feedback on the tools and the falsification approach: https://disconfirmai.com/en