Bixoto/decision-index-results

Decision Index 0.3 — Results

What this is: the complete, packaged results of Jex 1.0 on the Decision Index 0.3 public suite — 140,178 requests, one result row each — plus its score files, ladder, receipts and disclosures.

Verify the scores; do not train on them. These are evaluation results. The underlying suite contains text from dozens of upstream datasets under their own terms, several research-only or non-commercial or non-redistributable, and is not redistributed here. Every submitted row is re-scored on review, and this dataset is what the maintainers' re-score points at.

Jex 1.0 is a frozen google/gemma-4-31B-it checkpoint (no head, no LoRA, no fine-tuning) served by a deterministic, batch-invariant inference engine, with four disclosed probability-preserving post-processing levers. It is the highest-public-index no-training entry on the Decision Index 0.3 board.

---

Headline score

field	value
decision index (public index)	60.29
balanced skill	60.29
balanced raw	69.88
breadth skill	59.25
requests	140,178 / 140,178 ok
unsupported · error	0 · 0
complete	true
edition	0.3

The public index covers the 37-benchmark public panel (5 areas, chance-corrected, coverage-adjusted). It is 20% of the board's Full score; the other 80% comes from private tests only the maintainers can run. No Full-score or rank is claimed here.

The ladder that produced it

Cumulative, on the same 140,178 requests. The four levers add +0.81; each is reversible with one flag.

step	stack	public index
base	frozen capture, no levers	59.48
+ fold	cat62 fold (D, τ = −13.63)	59.96
+ c05best	cat59 isotonic · shrink λ = 0.75, nf = 10	60.09
+ R42	cat42 Entailment (τ = −8)	60.21
+ R20	cat20 option_0 (τ = +15)	60.29

---

Short story

Jex began as a production experiment at Bixoto: how far can a decision engine go without touching the weights? The board has exactly two serious frozen entries, and they run the same checkpoint — deck31b at 58.19 public, Jex at 60.29. The difference is entirely the engine around the model: the readout, the batching, the decision rules, and four disclosed post-processing levers. This dataset is that result, published whole — complete, byte-reproducible, and receipted.

---

Files

Everything lives under one run directory, runs/jex-1.0/:

file	purpose
results-combined.jsonl.gz	the run — 140,178 result rows (kit format); sha256 5003231c…ab9c7 (~88 MB)
scores.json	the submission file — decision_index 60.29, complete: true, counts, edition, suite hashes
index.json	per-benchmark and per-area scores, chance levels, coverage
benchmark-summary.json	native metric per benchmark, answered/unsupported/error counts
environment.json	engine options, pins, GPU/torch, latency definition
status.json	last runner event; {ok: 140178, unsupported: 0, error: 0}
combined-fold-receipt.json	the overlay receipt — no_probability_changes: true, per-lever counters
ladder-results.json	cumulative + leave-one-out results
receipt_diff.json	per-catalog movement — n_moved = 4, all others bit-identical
LADDER.md	the ladder, leave-one-outs, per-catalog deltas, rollback rule
REPRO.md	full reproduction chain + exact re-score command + frozen suite hashes
REPRO-INPUTS.md	reproduction inputs: what ships where, what is documented-only
INPUTS.md	exact construction & provenance of the reproduction inputs
DISCLOSURES.md	mandatory disclosures (lever provenance, fitting, determinism, rollback)
inputs/wave1-v1x-results.jsonl.gz	frozen v1.x base capture (140,178 rows; scores 59.48); sha256 59f27834… (~89 MB)
inputs/table.pkl	fitted calibration table (derived probability panel; 312,320 records); sha256 cf241d59… (~70 MB)
inputs/results-folded.jsonl.gz	cat62 folded capture (base + fold overlay only); sha256 5c9b96ae… (~87 MB)

---

Engine adapter & harness

• The rows were served by our own engine, Bixoto/jex (engine/, harness/, levers/) — a vLLM pipeline writing kit-format result rows, not the kit's stock http/transformers engines.
• Truncation policy: none. A prompt that does not fit its worker raises Unsupported and is counted as wrong; nothing is ever cut. Prompts above 24,576 tokens are declared unsupported (0 on the public suite).
• Option filtering: none. Every option in criteria carries a probability.
• Retries: the runner resumes from the results file; rows with status error are retried, all others are kept. The final run has 0 errors.
• Readout: a single greedy step over a 255-entry single-token label table (rendering reference, slot paren).

---

Hardware & latency

• Hardware: 1 × NVIDIA RTX PRO 6000 96 GB (Blackwell, sm_120), bf16, tensor parallel 1, serial (one request at a time).
• Latency (ours, n = 750, the board's own protocol): median 116.4 ms, mean 369.6 ms, p80 319.9 ms — all under the board's 1000 ms admission gate. p95 (≈ 1,745 ms) is displayed by the tooling but is not a gate.

• The maintainers measure latency themselves, single-process on one RTX PRO 6000, on a private held-out sample; that measurement — not ours — decides admission.

---

Reproducibility & verification

• Determinism: every path is a pure function of (state, questions) plus frozen constants — batch-invariant (VLLM_BATCH_INVARIANT=1) and serial (VLLM_ENABLE_V1_MULTIPROCESSING=0), greedy, determinism warm-up 3, TRITON_ATTN, FULL_DECODE_ONLY, tp = 1.
• Bit-exact re-score: the packaged results-combined.jsonl.gz reproduces decision_index 60.29 under two independent scorers — the kit03 CLI and the in-process score_run_v02 path.
• No-degrade: exactly four catalogs move (62, 59, 42, 20); every other catalog is bit-identical — the 43/43 no-degrade receipt.

pin	value
base model	google/gemma-4-31B-it @ 842da3794eaa0b77d5f08bae87a17459d91ff475
engine	Bixoto/jex @ 88dafa36ff31b828cc305f2cc996fcb88003dab3
kit / edition	62d2f51de34a2de64906345b6bc3e98e27ff55c7 (edition 0.3)
runtime	vLLM 0.30.0, transformers 5.14.1
results sha256	5003231c9757d430ff9439c17a0e20f166b8d3dacef1dcde66e49056708ab9c7
suite hashes	selected b2b56d6f… · added 7429f3c9… · gsm8k 75933489…
dataset revision	7efebd4c964459ec42b6de881cb6ac4871939aca

Re-score command (needs a checkout of the kit at 62d2f51 and the suite it builds):

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

---

Don't train on this

This repository holds results, not a training set. The Decision Index suite itself is evaluation data under per-upstream terms: several benchmarks are research-only or non-commercial (e.g. ANLI CC BY-NC 4.0), some forbid redistribution (RAGTruth's Yelp contexts), some are share-alike, GPQA asks that its questions not be posted in plain text, and HLE is gated. Do not train on the suite, and do not republish it.

---

Disclosures

• Public ≠ Full. 60.29 is a public index. The Full score adds 50% private same-skills and 30% private new-domains 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. combined-fold-receipt.json records no_probability_changes: true.
• Rollback: drop any lever by re-running run_variants.py with that flag False; the whole stack off returns to base 59.48.
• No #1 claim. The trained leader sits above the frozen ceiling; raising it needs training, which is out of scope by design.
• All work is AI-assisted.

---

Limitations & intended use

• Intended use: independent verification of the submitted Jex 1.0 run, and reference material for reproducing it. Not for training, not as a leaderboard claim, and not for redistribution of the suite.
• Capacity limits: choice (2–255 options) and noul only; bf16 (no quantization); prompts above 24,576 tokens are declared unsupported, never truncated.
• Private legs are maintainer-run. Small transport drift (±0.5) on the private parts is possible and not ours to control.
• The board's own table remains the source of truth for ranks and the Full score.

---

Links

• Engine repository — https://github.com/Bixoto/jex
• Model card (HF) — https://huggingface.co/Bixoto/jex
• Submission PR — https://github.com/apolinario/decision-index/pull/153
• Decision Index kit — https://github.com/apolinario/decision-index
• Base model — https://huggingface.co/google/gemma-4-31B-it

Citation

@misc{jex1_2026,
  title        = {Jex 1.0: a frozen decision engine for Decision Index 0.3},
  author       = {Bixoto},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Bixoto/decision-index-results}},
  note         = {Run jex-1.0; public index 60.29; complete 140,178/140,178 requests.}
}

_{Built at Bixoto for production use. Now open. · All AI-assisted.}