mini-jev

Quadrant scores

See the full quadrant

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

• Maturity2/10

  Preregistered teaching repo with demos — explicitly not a shipped decision product; lowest maturity on the map.

• Capability3.2/10

  Measures letter-logit reads for closed choices on frozen Qwen3-4B; catalog types are choice+boolean only (score not studied).

• Adoption7/100

  Useful citation in the method lane of S1Bench gaps; small audience beyond people studying select-over-generate.

Vendor claims

• On CLINC150 paired trials, letter-read approx equals JSON-grammar accuracy for closed enums (author preregistered study)[Vendor claim — not independently verified]
• Frozen Qwen3-4B-Instruct; no decision-head training (author)[Vendor claim — not independently verified]
• Score / ordered scales were not measured in the study (author)[Vendor claim — not independently verified]

mini-jev is a teaching experiment, not a trained System One weight release: it shows how far a frozen Qwen3-4B gets by reading letter logits for closed-choice fields versus grammar-constrained JSON — by r-ms, not TypeSafe.

Catalog decisionTypes are choice and boolean only — the study did not measure Jev-style score. Repo: github.com/r-ms/mini-jev.

Limits (read before you ship)

• Not a trained decision model — no RLCD, no pointer head, no released adapter claiming Jev parity.
• Letter shares are normalized candidate scores, not calibrated probabilities.
• Out-of-scope / abstention and dependent fields favour JSON generation in the author’s own results.
• Do not put this on a procurement shortlist next to Jev or Decider.

Why it is in the catalog

Closes an S1Bench documentation gap: the “select-over-generate on a frozen LM” method, clearly labelled so the quadrant does not confuse it with trained peers.