Methodology
The research is produced by AI-Inflect, a research system built for one question: is a company's business turning, and what is the evidence? It works in fixed stages, and each stage hands the next one a recorded, dated input. The short version is on How it works; this page explains each stage.
1. Company filings, as they were known at the time
Published reports are built from companies' own filings with the U.S. Securities and Exchange Commission: annual reports (10-K), quarterly reports (10-Q) and current reports such as earnings releases (8-K). Every figure is stored with the moment it became public. When a company restates a number, the new value is kept as a new version; the old one is not overwritten. This lets any result be reproduced with only the information that was available on its date, never with hindsight.
2. A deterministic screen
Eight quarters of reported results are turned into simple measures: revenue growth and whether it is accelerating, gross and operating margins, earnings per share and free cash flow. A fixed set of 17 fundamental conditions is then checked, such as “revenue growth improving” or “free cash flow margin expanding”. Each condition is true, false or unknown; missing data stays unknown and is never counted as a pass or a fail.
The screen is rule-based: the same data always gives the same result. Its rules are versioned and frozen. A change to them is tested against the previous version before it is adopted. No AI model takes part in this stage, and its output decides only which companies deserve a closer look.
3. Evidence from the filings
For companies the screen selects, an AI model reads each new filing and proposes factual statements, each with the exact passage it comes from. A statement is kept only if that passage matches the filing text character for character at the stated location; anything else is rejected. The model can miss things, so the absence of a statement is never treated as evidence that something did not happen.
4. Research and its counter-case
The verified evidence, together with the reported figures, is frozen into a dated research file. An AI model then drafts the analysis from that file only: what changed, the testable hypotheses, the risks, and for each hypothesis what would show it is wrong. When a new filing arrives, the research is redone and compared with the previous version, so changes in view are visible rather than silent.
5. Editorial review and publication
Before publication, every figure in a report is traced back to a numbered source, figures we calculate are marked as calculated, and the draft is reviewed. Published reports keep their history: corrections are marked, and earlier conclusions stay visible.
Model research notes are published as they are, clearly labelled as unreviewed AI interpretation; reports are reviewed. The notes (under AI-Inflect, Latest research) leave out the model’s suggested actions and any evidence comparing earnings with analyst estimates, and they never change a company’s stage.
What this method does not claim
- The screen finds companies worth studying. It is not a forecast, and no performance record is claimed for it here.
- AI output is interpretation that can be wrong; it never replaces the source documents, which every report links to.
- Research is general information, not a recommendation to buy or sell and not personalised financial advice.
The rules behind this are set out in our editorial principles.