morrenhale/decision-benchmark-jev-laya-julia

Decision-Model Benchmark 2026-09-27: jev vs laya vs Julia-1-ONNX

Raw data behind the head-to-head benchmark of three decision engines: 19,776 decisions × 3 engines (59,328 inferences), 0 inference failures. Everything needed to audit or re-score the run.

Contents

• manifests/*.jsonl — frozen test manifests (19,776 decisions): question wording, criteria, option order (keys), gold key, per-dataset row ids. manifests/meta.json records source URLs and SHA-256 of every dataset file (typed-decisions pinned rev c76749ec…).
• predictions/{jev,laya,julia-onnx}.jsonl — 19,776 rows each: predicted key, full probability vector over keys, gold, latency_ms, engine-specific traces (julia tournament rounds/final candidates; jev input_tokens).
• predictions/*.latency.jsonl — 250 rows per engine from dedicated serial runs (one engine at a time, jev single in-flight). The only latency numbers reported publicly come from these files.
• results.json — all computed metrics: accuracy, macro-F1, Brier, clipped NLL, ECE, p(gold)=0, per-dataset and per-slice breakdowns, paired-bootstrap deltas (20,000 resamples, target-stratified, seed 20260927).

Test sets (gold comes from the datasets themselves)

Dataset	Decisions	Source
typed-decisions	2,000	LocalLLaMA/typed-decisions test parquet (official TypeSafe/Julia suite: 600 choice, 800 score, 600 noul)
AG News	7,600	full test split
DAIR Emotion	2,000	full test split
Banking77	3,076	full test split (77 options per call)
MASSIVE scenario	5,100	51 locales × 100 rows, deterministic SHA256-based selection, seed 20260927

Question wording and criteria are identical across engines and frozen in the manifests.

Engines

• jev-1.13.0 — TypeSafe hosted API (POST /v1/systemone)
• laya 0.3.20 — NandhaKishorM/laya shipped Router, base zero-shot (the laya-typed-decisions fine-tuned checkpoint was deliberately excluded: train-on-test on this suite)
• Julia-1-ONNX — SupersonicLabs/Julia-1-ONNX published model.onnx + model.onnx.data on onnxruntime CPU, native julia.data.sequence encoding (strict, max_length=1024, head_length=512);

  20 options via the official julia.router.Router tournament (width=20, survivors=2). ONNX↔torch parity verified: argmax 40/40, max |Δlogit| 0.000258.

Scoring conventions (matters for calibration metrics)

• jev returns probability maps with sum drift up to 0.01 (display rounding is the working hypothesis — see the run's error notes); 416 rows affected. laya probabilities pass through display rounding (round(4)). All vectors are renormalized before scoring.
• julia Banking77 probabilities are conditional on the final tournament candidate set (probability_scope=final_candidates) — the vendor's documented Router behavior.
• NLL is clipped; ECE is 15 equal-width bins.

Caveats

• MASSIVE uses bare label names as criteria (hard zero-shot). The Julia-1 card's 71.5% used scenario descriptions — not comparable to the 0.402 measured here.
• Reported latency is end-to-end deployment latency on different hardware (jev: hosted API + network RTT; laya: Apple MPS, torch 2.14; julia: onnxruntime CPU), not compute-normalized speed.
• Vendor-claim checks and the full write-up live in the companion report (BENCHMARK.md in the release zip).

License

Published dataset rows remain under their original dataset licenses (typed-decisions, AG News, DAIR Emotion, Banking77, MASSIVE). Prediction and manifest files are CC-BY-4.0.