How accurate are the odds?
The recipe behind Halomark's model: Kev, fine-tuned on 2,000 founders, reading the answers to TRL's 40 questions. These are its scores on 600 test founders it never learned from or was tuned on, for the yes/no question ($500M or more). Halomark's own version, trained the same way, also predicts the outcome band and timing.
When it gave a chance, how many succeeded?
Each dot is a group of founders given similar odds. On the dashed line, the odds were exactly right.
Show as a table
| Chance given | Average | Succeeded | Founders |
|---|---|---|---|
| 0%–10% | 7% | 9% | 99 |
| 10%–20% | 15% | 17% | 218 |
| 20%–30% | 25% | 25% | 238 |
| 30%–40% | 33% | 36% | 45 |
In plain words
- Typical gap between the odds and what happened
- 1 ptOn average, the chance it gave was within about 1 percentage point of the share who actually succeeded.
- Of the founders it rated highest, this many succeeded
- 3 in 1012 of the 36 it rated highest reached $500M+, against about 2 in 10 of all 600 founders.
- Of the founders who succeeded, the share it rated highest
- 10%It picked out 12 of the 120 who succeeded. Most successes look ordinary on paper, which is why a pass is never a no.
The honest limit: these are 600 founders from a mostly generated dataset (2,950 of 3,000 profiles are synthetic), so they show how honest the odds are on that data, not on real founders. It never gave a founder odds above 40%: it is cautious, so it rarely singles anyone out, and the right of the chart is empty.
More about the model and its data in the model card.
Source: results/summary/d1-final (C3, test) on the research branch. Updated 2026-11-15.
Screening aid, not investment advice. The model estimates the chance of a large outcome from a founder's background on the day the company was founded. It was trained on generated data, so its accuracy on real founders is unknown. Public information only; nothing is stored after the session.