Benchmark
Internal blind test, October 2026
Before an expert’s report is served, how much of it can be predicted? We tested that on real expert reports, blind, and changed only what the model was allowed to know. With the expert’s data, methods and record, Supreme Mind reproduced nearly three times as many of the expert’s key findings as the same model on its own.
The result
The score is the share of the expert’s key findings reproduced within tolerance. Each arm adds information to the one before, except E, which adds the news timeline to B alone.
+50 points, A to F80% interval: +37 to +60 points
Case data and the ability to run the analysis account for most of the gain. Knowing the class of expert, and then the individual expert, adds to it, and Supreme Mind, with everything, does best.
What each arm was given
- A. The model on its own: the case file and nothing else.
- B. Case data and code execution: the market data the expert would use, and the ability to run the expert’s analysis on it.
- C. Expert-class profile: the methods and choices typical of this kind of expert, drawn from public expert filings, with no individual named.
- D. Individual expert dossier: the individual expert’s public prior work.
- E. Dated news timeline: dated news and company press releases for every trading day, added to B.
- F. Supreme Mind: all of the above, together.
What each layer adds
Data drives the findings. The individual dossier drives the method. Each row is the change from one arm to the next, on two measures: the expert’s key findings, and the main method choices the expert made.
On the findings, Supreme Mind gains 50 points over the model alone, and case data and code add 34; both clear zero even at 95%. The class profile and the news timeline add smaller gains. On method, data alone slightly lowers the matches and the individual dossier restores them, with news on top of the data adding to both.
Questions about the benchmark? Email richard@suprememind.ai.