Did EnemoAi see it coming?

The companies

Two groups. Failures: public companies that filed for bankruptcy, had their bank placed in receivership, or wound down their main business. Survivors: companies that showed real distress (a going-concern warning, an emergency cash raise, a run on deposits, a default notice) and did not fail in the next 24 months. The survivors are there so false alarms are counted, not only hits. Every date and outcome comes from the company's own SEC filing or the FDIC, linked on each company.

The replay dates

Failures are replayed 24, 12, 6 and 3 months before the event. Survivors on the day their distress was disclosed and six months later. On each date EnemoAi sees only filings and records dated on or before that day.

What is scored

• Live Threat Alerts: the alert grading EnemoAi uses today, applied to the company's filings. A company counts as flagged when a critical or high alert fired in the 12 months before the replay date. No AI is involved, so it cannot know the ending.
• Mentor: the question a supplier or lender would ask, answered as of the replay date, ending with a stated chance of failure within 24 months. 50% or more counts as a warning.
• Fate Radar, when it is built, is scored the same way.

No peeking

Every passage shown to the model and every source it cited is checked against the replay date. One source dated later, or undated, voids the whole run; it is not published. One limit we cannot test away: the model behind Mentor was trained on text written after these dates and may remember how the story ended. It is told not to use that, and every source is checked, but its memory can't be audited. That is why the alert replay (no AI) and the forward calls (made before the outcome existed) are shown separately: they are the cleanest tests.

Reading the numbers

• Hit rate: of the failure replays, the share flagged. False-alarm rate: of the survivor replays, the share flagged.
• Lead time: for each failure, days from the first warning to the event.
• Calibration (Brier score): how close stated chances were to what happened, from 0 (perfect) upward; always saying 50% scores 0.25. It is compared with always stating this set's own failure share.
• Every number shows how many cases it rests on and a 95% range. The ranges are wide because the set is small; treat them as such.
• This set has more failures than real life, so "how many warnings were right" here is higher than it would be across all companies. Hit rate and false-alarm rate don't depend on that mix.

The full method, what isn't measured yet and where things stand today are on the methodology page (in English).