I tried TypeSafe's System One model

They are vendor launch figures restated in the description. Nothing in the video or its repo produces them, so if the multiple matters to your decision you would need to reproduce it against your own traffic. They also assume no output-token fee, which is the main thing that drives them.

The framing is that Jev handles routing, filtering and pre-processing so the generative model is only called on the cases that need it. That reduces calls and tokens on a filtering job, but it also means a wrong filter silently drops work before the better model ever sees it.

Demo code is linked at a public repository, so you can. Nothing documents its dependencies, how it is configured, or whether it ships with the transcript data it analyzed, so expect to supply your own input before you get comparable numbers.

The video does not cover it. The failure mode of a cheap pre-filter is silent under-inclusion rather than a visible error, and the linked demo gives you nothing to build a fallback or a review queue from without measuring your own false-negative rate.

Joe Maddalone published it, with demo code linked at the joemaddalone/jev-transcript-analyzer-demo repository.

As a typed probability engine for routing, filtering, and pre-processing before a generative model, so the LLM is only called on cases that need it.

No. The entry says the 193x and 444x figures are vendor launch claims restated in the description, not a measurement produced by this video or its repository.