Three previews of a MacBook-sized Jev rebuild miss the release screen their author declared

A trained layer produces the number. A pointer head scores each question's options against its decision token, applies a softmax, and those probabilities are the answer ^[4]. The head learned them with cross-entropy against labelled outcomes, which the README describes as probabilities "learned rather than generated as text" ^[5]. Three question types go through one shared readout: yes/no, a choice of 2 to 255 options, and ordered score levels ^[7]. The design follows a published reconstruction of TypeSafe's Jev, "Jev's Architecture Unmasked" ^[8].

Training against outcomes fixes the shape of the output. Whether the numbers hold on a document unlike the training set is a separate question. On the out-of-domain suite, kev-8b posts a Brier score of 0.34 and 8 percent confident errors on sources it never trained on ^[10]. That rate needs a denominator. If it is the 764 out-of-domain records, about 61 came back confident and wrong ^[26], and an auto-escalation rule would have acted on all of them.

Pack many questions into one sequence and one specific failure comes with it: a later question conditions on an earlier one's wording. The block-causal mask is the guard ^[3]. The evidence it worked is that packed requests and separate requests agree to 4e-6 ^[6]. That agreement is what rules out sibling leakage as an explanation for the parallel scores.

The table's rows all sit on the same frozen, checksummed items: 1,200 in-distribution questions from the trained sources, and 764 out-of-domain records from six public sources plus held-out programmatic policy rules ^[13], for 1,964 items in total ^[22]. Doubling from 4B to 8B parameters moved the out-of-domain score by 0.01 ^[24]. Jev sits 0.09 above kev-8b ^[23]. The README gives these as bare numbers and does not name the metric.

The items are matched on difficulty. Serving kev-4b costs a 32 GB Mac running bf16 ^[16]; a Jev call costs whatever TypeSafe charges for it.

The release screen is the part I would want other small-model releases to copy. The gate was set in advance: both siblings of a policy pair correct at 70 percent or better on held-out rule reasoning. The best preview read 0.67, three points short ^[14][25]. Each model card records one locked-test read, so no best-of-several selection sits inside the figure ^[14].

Serving is where the laptop comes in. kev-0.5b trains in about an hour and three quarters on an Apple M5 ^[15], while the 4B and 8B recipes take 40 to 70 minutes on a single H100 rented through Modal before they serve on the Mac ^[16]. Serving is tested on Apple Silicon; training and evaluation ran on CUDA and MPS ^[18]. The README points new users at kev-4b, "the best accuracy per byte" ^[17]. In its own worked example, that checkpoint answered a choice, a yes/no and a score question about one customer complaint in 277 milliseconds on an M5 ^[19], and the yes/no on whether the case needed urgent human attention came back at 0.54 ^[27].