Jev by TypeSafe AI — Interactive Talk & Pro Research Dossier

Jev by TypeSafe AI

SYSTEM 1 MODEL

Jev Research & Benchmark Workbench

Synthesizing findings from all 8 community & technical sources

Archer Hume 10,000 API-Probe Study

Jev’s Architecture Unmasked: 8 Components

Source: archerhume.com/posts/jevs-architecture-unmasked

The Proposed Computational Graph

Prefix Shared State

KV cached once. Up to 32,768 tokens. Processed across all questions simultaneously.

Question Tree Suffixes

Isolated question branches. Tree attention mask prevents cross-question leakage.

Linear Readout Head

z = Wh + b + softmax/sigmoid. Zero autoregressive decode steps.

1. Causal Decoder vs Encoder

Jev acts encoder-like, but Hume concludes it is a causal transformer decoder (likely sparse MoE with ~10B active parameters out of 70B+). 30k tokens were processed in ~160ms. MMLU-Pro accuracy is 84.6%, indicating massive pretraining weights.

2. Listwise Option Cross-Attention

When an irrelevant 5th option ("weather caused it") was added to a 4-option set, the log-odds between existing options shifted by -0.28 across 10 randomized blocks. This proves options attend to each other, resolving contextual terms like "none of the above".

3. Context Limits & Token Accounting

Max branch length: 32,768 tokens (2¹⁵).
Max total packed request: 65,536 tokens (2¹⁶).
State is counted once: a 23k-token state with 1,000 questions fits in one call without duplicating context.

4. RLCD: Proper Scoring Rule

Trained via proper scoring rules (Brier loss & log-loss). On a 1,200-item MMLU sample, Jev achieved an Expected Calibration Error (ECE) of 0.0313, fixing standard LLM overconfidence.

Benchmark Heaven (JevBench)

Jev-Class Model Leaderboard

Source: benchmarkheaven.com/jev-models · Evaluated on 4 equal 25% axes

Rank	Model / System	Base Engine	Overall	Intelligence	Calibration	Latency	Cost / 1k
#1	decider-4b v2 (Mapika)	4B Dense	64.1	49.4	75.0	0.02 s (20ms)	$0.020
#2	Jev 1.13.0 API (TypeSafe)	MoE (~10B active)	63.3	53.1 (Top)	76.3 (Top)	0.65 s (650ms)	$0.040
#7	JEV Qwen3.5-9B NVFP4	Qwen3.5-9B	57.3	46.8	67.7	0.19 s	$0.077
#8	Winnow-12B Q8	12B Q8	56.6	48.3	64.8	0.60 s	$0.037
#16	Kev-4B (Jared Palmer)	Qwen3.5-4B	52.3	43.1	61.5	0.33 s (330ms)	$0.019
#41	Laya (ConvAI)	ModernBERT 421M	49.9	36.1	63.7	0.033 s (33ms)	$0.0029 (0.07x)