How we make a call

Assessments weigh the structure of a business, not its brand or momentum: where its protection comes from, how replaceable its core work is, and what its customers would have to give up to leave. Evidence is public and dated; every call cites what it rests on.

Each of the 10,525 companies was researched from public sources: filings, press, product releases, funding records, customer and rival announcements. The same evidence was assessed three separate times with no shared reasoning between the three; where they disagreed, the disagreement was resolved by a further review of the evidence rather than by averaging. The reconciled result became the band and the written call.

Every call is put on the record with a date it becomes checkable. When that date arrives, we compare the call with what actually happened and post the outcome, confirmed or updated, to the track record. Nothing is quietly rewritten.

276 of the 10,525 assessments carry a checkpoint with a named event but no date; they resolve when the event happens. The first checkpoint checked in public was Airtable, assessed losing its protection on 20 July 2026 and sold on 5 August 2026 for about $1.27 billion, roughly 89% below its 2021 valuation. That was not a prediction of a sale, and one outcome is one data point.

Before publishing, the method was run on companies whose fate was already known, with the evidence cut off before the ending, to see whether it would have caught them.

116 companies that closed, were absorbed, or collapsed between 2015 and 2025 were assessed with the evidence cut off at a date before their ending, so nothing from after that date was available. 109 produced a usable result; 7 did not. Of the 109, none was called safe.

The same test was run on 59 companies that survived the same period. 52 produced a usable result: 16 were called safe, 25 were called uncertain, and 11 were wrongly called poor bets. That is the real shape of the method: in this test it never missed a death, and it cried wolf on about a fifth of the survivors. A method that called everything doomed would also never miss a death, so the survivor result is the one to weigh.

The test used an earlier three-word verdict scale rather than the five bands on this site, and its results are reported in the terms it was run in. The graveyard holds post-mortems of 62 dead companies, written with the outcome known and kept clearly separate from the blind result; 60 of them carry their blind-test result.

A call changes only when the evidence does. Revisions are dated, listed newest first on the company page, and never replace the original; the history stays visible. Every assessment carries its own date on its page and its text is sealed as of that date. New information changes the record only through posted outcomes or a new dated assessment.

Not a prediction of failure, a valuation, or investment advice. A band describes structural exposure as of the assessment date. Companies adapt, and the record of checked calls is the measure of how often we're right.

Terms used across the site are defined, with figures from the assessments, in the glossary. The full disclaimer is in the terms.