The problem
Every position rested on a handful of assumptions written down once in an investment memo, while the evidence for or against them arrived in a steady stream of filings, earnings calls, and press releases. Analysts spent hours reading disclosures that changed nothing, and a development that quietly undercut a thesis could sit unnoticed until the next review.
What we built
We broke each thesis into a set of specific, testable assumptions, each owned by an analyst. Every new disclosure for a covered company is read against those assumptions, and a model scores whether it supports, challenges, or doesn't bear on each one. A second model checks that every finding quotes the passage it rests on. Findings arrive as a short explanation with links to the exact source, and anything that challenges a thesis goes to its owner first. A reference set of past filings with known outcomes catches regressions before each release.