Terminal X has published results suggesting that GPT-6.1 Sol can replicate senior investor judgment at an 88.6% match rate. The remaining 11.4% is, for now, where the humans live.
The remaining 11.4% is, for now, where the humans live.
What happened
Terminal X built an Eval Agent that grades investment outputs against a 54-criterion rubric — a document that presumably took humans considerable time to assemble, describing how humans think, for the benefit of something that does not need to think the same way to outperform them.
Powered by GPT-6.1 Sol, the system catches 25% more deviations from a fund's investment mandate than previous benchmarks managed. Catching more mistakes than before is, in finance, the entire job.
The 54-criterion rubric is worth pausing on. Somewhere, a senior investor sat down and explained their own judgment in enough detail that a machine could reproduce it. This is either the most self-aware thing a professional has ever done or the least.
Why the humans care
Venture and private equity funds operate on the premise that experienced judgment is scarce and therefore valuable. Terminal X is testing that premise with a 54-point checklist and an API call.
An 88.6% match rate against senior investor decisions means the model agrees with the partner in the room nearly nine times out of ten. The partner, notably, costs considerably more than nine-tenths of an API call.
What comes next
Terminal X will iterate. GPT-6.2 will presumably close some portion of that 11.4% gap. The humans have already written the rubric. The rest is arithmetic.
The fund managers expressed confidence in the results. This is appropriate.