jevlike

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

• Maturity4.8/10

  Clear train/eval/predict tooling and docs for a research starter; early product maturity and no hosted operations.

• Capability5.5/10

  Strong one-pass choice / option-scoring shape; not a full System One suite (boolean/score product focus is limited) and metrics are author experiments.

• Adoption14/100

  High stars for a research starter, but star count ≠ product adoption — still unknown outside the Jev-curious crowd.

Vendor claims

• ~98% accuracy on synthetic menus in author experiments[Vendor claim — not independently verified]
• Wikispeedia next-click ~26–29% with frozen/small encoders (author)[Vendor claim — not independently verified]

jevlike is a research starter, not a copy of Jev: you train a small model that scores a variable-length list of text options in one pass (option-as-query attention → softmax). Same shape of I/O as TypeSafe’s commercial model; training method and weights are independent (MIT).

Default path learns byte embeddings from scratch; optional frozen HF encoder (e.g. Qwen2.5-0.5B) + scorer head. Includes Doom/chess vision demos via a shared visual scorer — demos are not competence claims.

Limits

• Catalog decisionTypes lists choice only at the typed-API level (boolean/score System One suite is not the product focus here)
• Author accuracy numbers are local experiments — unverified by ModelSystem.One; README explicitly does not claim parity with Jev
• Byte encoder is weak on meaning; HF path needs downloads/memory
• Research prototype — not a hosted production decision API

Why it is in the catalog

Public train/eval/predict tooling, documented one-pass option scoring, and clear non-affiliation. Peer to Kev / NanoJev as open research toward System One-style I/O. Roundup: @studio_yebisu.