Results
What Diaspore has shown so far. Every result sits beside what it has to beat: chance, a simple rule, a hand-set scorecard, and an AI given the same facts.
Each test pretends it is an earlier date, uses only what was public that day, and is checked against what happened afterwards. The test and its pass mark were written down before the result was read.
Young companies
Which new companies will raise a seed round?
8 of the top 25 picks made on 2022-12-31 raised within 24 months.
| Picking at random | 1.7 |
| The best simple rule | 5 |
| A hand-set scorecard | 6 |
| An AI given the same facts | 7 |
A fresh date: the model was fixed first and scored once. Over the whole list the model and the AI are level.
Labs
Does it find the labs that fame misses?
27% of the 190 labs that later formed a funded company, and that a ranking by citations had left out of its top tenth, were in the model's top tenth.
| By chance | 10% |
| Ranking by paper count | 9% |
| Ranking by research dollars | 13% |
| A hand-set scorecard | 18% |
The model was designed with this date in view, so read this as a best case, not a fresh test.
Labs
Which labs will form a funded company?
74% of the time the model ranked a lab that went on to form a funded company above one that did not, on two past dates, trained each time only on outcomes already known.
| A coin flip | 50% |
| Ranking by paper count | 66% |
| A hand-set scorecard | 69% |
The gap over paper count is 9 points (between 6 and 12). The model was designed with these dates in view, so read this as a best case, not a fresh test.
Reports
Does a report say what its sources say?
161 of 164 claims in five reports on real companies at past dates matched their source when a second AI, which had not seen the runs, checked each one.
| Wrong claims slipped in on purpose, and caught | 24 of 24 |
This checks that a report is faithful to its sources. It does not check that the sources are right.
Methods, comparisons and known limits: how it was tested.