Automated README summary
Automatically summarized from the README on 20 Sept 2026; not a run by jevbooks. Run instructions and author-reported outcomes still need verification.
What it is
A preregistered experiment and teaching bench measuring whether a frozen Qwen3-4B can answer closed-choice schema fields by reading next-token logits for option letters, instead of generating grammar-constrained JSON. Task: CLINC150 intent classification.
How it uses Jev
Each closed-choice field becomes a lettered multiple-choice question; one forward pass reads the model's scores for the option letters at the answer position. No token is generated for choices. Strings and numbers are still generated under a grammar. Results are compared against JSON generation.
Primitives:choicenoul
Technique worth stealing
Give the model a one-token identifier (a letter) to answer with, then read the logits instead of decoding.
Try it
uv sync; MINIJEV_DEVICE=mps uv run python demo/server.py (or cuda), then open http://127.0.0.1:8765/.
Model judgments from README excerpts
Project signals use repository metadata, an existing summary and the first 700 README characters, not the full project. “Not found” means not detected in that material; it does not establish absence. These are the original judgments with recorded models and dates, not source audits or run results.
Models & evaluation dates
- Inclusion gate
jev-1.13.0- Project signals
jev-1.13.0- Category
jev-1.13.0- Pattern tags
jev-1.13.0
Evidence
These signals use repository metadata, an existing summary and at most 700 cleaned README characters. ≥ 0.60 means reported in that excerpt; ≤ 0.40 means not found there. The middle band is inconclusive. None is a source-code audit or a test result. The runnable-tool question excludes research experiments; the recommendation question asks for a learning shortlist, not a quality verdict.
- Jev is core to the projectFound in material0.81
- Shows a System One patternunclear0.41
- Describes handling uncertaintynot found in the excerpt0.07
- Reports measured numbersnot found in the excerpt0.18
- Is a runnable toolnot found in the excerpt0.12
- Top-10 learning shortlist signalnot found in the excerpt0.22
- Is a replica of the modelFound in material0.93
- Solves a common problem (0–2)score on a 0–2 scale0.74 / 2
- About JevFound in material0.97
no-judgment band (0.40–0.60)
no-judgment band (0.40–0.60)
The Jev model and date for each judgment stage are listed above. Card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.