Jev's Architecture Unmasked

No, and that is one of the sharper findings. Token counts fit a pattern of about four shared tokens plus fifteen per answer, and the field is derived from the length of the returned string after the fact for billing. There is no generation loop at all; inference ends right after prompt processing.

Through latency scaling and token accounting. Input tokens are additive across questions, 268 for one yes/no question and 276 for two, which implies shared context is cached once rather than recomputed, and a 23k-token state with 5,000 questions still fits the implied ~65,536-token ceiling. Those are measurements, not guesses.

Two tiers, and the author is explicit about them. Shared computation and direct probability readouts are well supported by the measurements; KV sharing mechanics, the exact readout type and especially the sparse mixture-of-experts backbone are progressively more speculative. Nothing here is confirmed by the vendor.

It can flip your decision. Reversing the order shifted probabilities from the 0.84-0.89 range to 0.93-0.96, and the options interact: adding an irrelevant distractor changed the relative odds between the originals. The throughput case is real too, since 1,500 questions came back at a 610 ms median.

That is the design claim. In the schematic routing request the queue choice assigns payments 0.91 while the escalate noul sits at 0.42, and the argument is that those are different uncertainties, so software can route a ticket automatically on one and leave escalation to a separate policy. The post is explicit that the probabilities in that example are invented.

Roughly 10,000 API calls, published 17 September 2026, hunting for signatures in latency scaling under different context lengths and question reordering, plus public documentation, research and prior art. Everything here targets jev-1.13.0 specifically, and the post carries a section on what would change the author's mind.

From the readout rather than the output. The argument is that a pretrained LLM generates "90 percent confident" as text, and the probability of producing those words establishes nothing about the probability of being right. Jev's proposed move is to keep the model's knowledge and read decision probabilities straight out of internal representations trained against outcomes.