Jex 1.0

This repository contains no weights.

Jex 1.0 serves the frozen google/gemma-4-31B-it checkpoint at revision 842da3794eaa0b77d5f08bae87a17459d91ff475 (bf16) through our own deterministic, batch-invariant inference engine. The weights are not copied here: fetch them from google/gemma-4-31B-it at that revision. We did not train, fine-tune, quantize, or add a head or LoRA — the checkpoint is untouched, and the whole difference is the engine around it.

Jex 1.0 is a decision engine, not a language model. Given a state and a set of typed questions, it returns one answer per question with a probability for every allowed option — in a single forward pass, no free-form text, no output parsing. It cannot answer with something that was not on the list.

Headline

field	value
Decision Index 0.3 public index	60.29
balanced skill / raw	60.29 / 69.88
breadth skill	59.25
requests	140,178 / 140,178 ok
unsupported · error	0 · 0
complete	true
base model	google/gemma-4-31B-it @ 842da3794eaa0b77d5f08bae87a17459d91ff475 (bf16, frozen)

The board's next-best frozen entry, deck31b, runs the same checkpoint and lands at 58.19 public. Jex 1.0 lands at 60.29 — a +2.10 difference that is entirely engine work: the readout, the batching, the decision rules, and four disclosed deterministic post-processing levers.

60.29 is a public index — the chance-corrected score over the 37-benchmark public panel (5 areas). It is not the board's Full score, which adds private tests only the maintainers can run. No Full-score or rank is claimed here.

Quick links

• Engine repository — https://github.com/Bixoto/jex
• Results dataset — https://huggingface.co/datasets/Bixoto/decision-index-results
• Submission PR — https://github.com/apolinario/decision-index/pull/153
• Base model — https://huggingface.co/google/gemma-4-31B-it

How to run

The engine runs on vLLM and needs one RTX PRO 6000-class GPU (96 GB, sm_120). Setup, run and re-score instructions live in the engine repo's README Quickstart: github.com/Bixoto/jex. The pinned configuration:

setting	value
runtime	vLLM 0.30.0, transformers 5.14.1
batch invariance	VLLM_BATCH_INVARIANT=1
multiprocessing	VLLM_ENABLE_V1_MULTIPROCESSING=0
concurrency	serial — one request at a time
sampling	greedy, determinism warm-up 3
attention backend	TRITON_ATTN
cudagraph mode	FULL_DECODE_ONLY
tensor parallel	1
readout	single greedy step over a 255-entry single-token label table (rendering reference, slot paren)
temperature	per option count: max(0.05, 10.124 − 1.633·ln n)
router	single_long: ≤ 8190 → 8192 worker · longer → 24576 worker · > 24576 → unsupported

# re-score the submitted results — reproduces 60.29 bit-exactly
PYTHONPATH=<kit03> python -m decision_index.cli score \
  --edition 0.3 --suite-dir <suite03> \
  --results runs/jex-1.0/results-combined.jsonl.gz \
  --out runs/jex-1.0/score
# expect: decision_index 60.29 · complete true · completed 140178

<kit03>/<suite03> are a checkout of the Decision Index 0.3 kit at 62d2f51 and the suite it builds. The full chain and hashes are in the engine repo.

Disclosures (condensed)

• Public ≠ Full. 60.29 is a public index; the Full score adds 50% private same-skills tests and 30% private new-domain tasks run by the maintainers.
• Positional construction-reuse (fold/cat62, R20/cat20). Both key on the literal option position/identity the construction assigns to the gold answer (D; first option). Genuine on the public suite (cat62 skill 0.6904 → 0.8324; cat20 0.8613 → 0.9094), but unverifiable on a private run whose construction is not a twin — disclosed as public-suite projections.
• Public-OOF fits (c05best/cat59, R42/cat42). Selected on the public out-of-fold panel; a private panel may regress toward the base rate.
• Probabilities are never modified. The levers rewrite only the emitted choice (recovered to the pre-TBO rule space) and the cat59 noul field (no_probability_changes: true).
• Rollback. Drop any lever by re-running run_variants.py with that flag False; the whole stack off returns to the base 59.48. The ladder is 59.48 → 59.96 → 60.09 → 60.21 → 60.29.
• Determinism. Every path is a pure function of (state, questions) plus frozen constants — batch-invariant and serial. The packaged combined file re-scores bit-exact 60.29 under two independent scorers.
• No #1 claim. The trained leader sits above the frozen ceiling; raising it needs training, which is out of scope by design.

License

The engine code in Bixoto/jex is Apache-2.0 (© 2026 Bixoto). The base weights are not redistributed here and remain under Google's Gemma Terms of Use; the Decision Index suite is evaluation data under its own upstream terms.

---

_{Built at Bixoto for production use. Now open.}

All AI-assisted.