autotrust/jev-decision-index-results

JEV models on the Decision Index

Complete Decision Index 0.2.1 runs of the JEV typed-decision models by AutoTrust AI, for the Jev Decision Index board (kit). Only engine outputs are published here (compact result rows: run ids, statuses, answers and probabilities, timings). No benchmark inputs are included; rebuild the suite with the kit.

run	model	Decision Index 0.2.1	raw	complete
runs/jev-9b	autotrust/JEV-9B @ 4ab5dfb	43.14	56.91	yes (150,317 of 150,317 scoreable requests, 0 errors, 0 unsupported)
runs/jev-27b	autotrust/JEV-27B @ 9b1f6fd	53.30	64.32	yes (150,317 of 150,317 scoreable requests, 0 errors, 0 unsupported)
runs/jev-gemma4-26b-a4b	autotrust/JEV-Gemma4-26B-A4B @ 603bfed	58.05	67.35	yes (150,317 of 150,317 scoreable requests, 0 errors, 0 unsupported)

runs/jev-9b

file	contents
results.jsonl.gz	150,759 result rows in the kit's format (--compact: no payload / raw output); the 442 excluded questions were run as well
scores.json, index.json, benchmark-summary.json	output of the kit's score --edition 0.2.1 (re-scoring results.jsonl.gz reproduces them exactly)
environment-sequential.json, environment-batched.json, status.json	the environment records of the two phases of the run (below) and the final status
source/jev_engine.py	the engine (jev_engine:JevEngine), sha256 8ee479f6…
source/fast_run.py	the batched driver used for the second phase, sha256 14056c7f… (published with relative paths; otherwise the file that ran, plus the later, unused vLLM backend)
training-overlap-check.json	verbatim-overlap counts between the training corpus and the suite

Suite. Rebuilt locally with suite rebuild and imported with suite import: selected-rows uncompressed sha256 b2b56d6f… and added-rows 7429f3c9…, exclusions 331df32d…, all subset files matching. (RouterBench's normalizer averages floats with the built-in sum(), whose algorithm changed in Python 3.12; rebuilding it under 3.11 gives different state digits, so the suite was frozen under Python 3.12, which reproduces the pinned hashes.)

Engine. The model's own System 1 path, as on its model card: Qwen3.5-9B text backbone in bf16 (fp32 tensors kept fp32), the adapter/ LoRA merged in memory, the 24-slot fp32 decision head from head.safetensors, per-kind temperatures from calibration.json, one prefill pass per question, read-out at the last token of the model's bare-v1 template. On the model's own held-out test rows the engine matches the project's reference inference to max |Δp| 3.6e-07 with identical argmax (300 of 300). Mechanical translation, identical for every benchmark:

• state, instructions and option descriptions verbatim when strings, JSON otherwise; a choice option is rendered "<key>: <description>", or "<key>" when the description is empty or equal to the key;
• a noul statement is phrased "Is this scenario one where: <statement>?", the form in which all 150,515 TypeSafe Jev 1.13-labelled noul rows of the training corpus present a statement (already-questions are passed verbatim). This was fixed before any suite result was seen; it only affects RAGTruth among scored benchmarks.

Declared capacity. The decision head has 16 choice slots (A–P). A question with more than 16 options is answered with every option read by the model and none pruned, in ceil(n/16) + 1 passes: contiguous groups of near-equal size (≤ 16) in the given order, then a final of 16 (each group's top option, free places filled by in-group probability); finalists keep the final's distribution times the chance the answer is among them, every other option gets its group's share of the final times its in-group probability. Context: 262,144 tokens; nothing is truncated (a longer prompt would be refused; none was). Question types: choice and noul.

How the run was executed (two phases, same engine, same output file).

1. Rows 1–35,590 (BFCL through the start of POP909-CL) ran with the kit's own sequential runner (run, then pipeline resuming on the imported suite): environment-sequential.json. Every finished row's payload_sha256 was checked against the imported suite before resuming (0 mismatches).
2. Rows 35,591–150,759 ran with source/fast_run.py, which answers 64 requests at a time: the prompts of all 64 are rendered by the same engine code and share length-sorted forward passes (prompts over 16,384 tokens run alone; right padding without an attention mask, which is exact for this causal model because pad tokens follow the last real token). environment-batched.json. The script's vLLM backend was added after this run and was not used; its default (--backend hf) path is the one that ran.

Agreement of the two execution paths, measured by re-running 3,000 randomly chosen phase-1 requests through the batched driver: 7,044 of 7,051 answers identical (99.90 %); the 7 differences are near-ties (probability differences of 0.004–0.013). Because of the batching, total_wall_ms in phase-2 rows is the batch wall time divided by 64, not a per-request latency. Sequential single-request latency on the same GPU (compatibility pass, kit runner): median 77 ms.

Hardware. 1 × NVIDIA B200 (183 GB), torch 2.13.0 + CUDA 13.0, transformers 5.16.1, peft 0.21.0, kit 87d4650.

Training data and overlap. The model was trained only on the public SargeDev/jev-distill-corpus-v3 (streams yuri_v3, TypeSafe Jev 1.13 output distributions over synthetic operational scenarios; openjev_v2, Open-Jev synthetic games/workflow tasks; yuri_v1, memory-relevance pairs with placeholder labels). Checkpoint selection and temperature fitting used the corpus's own validation and calibration splits; no suite data was used. Check over the train, validation and calibration splits (683,683 rows) against every suite state, instruction and option: no whole state or question of the suite appears in the training data; zero sentence-level hits in yuri_v3 and openjev_v2. 547 yuri_v1 rows share Wikipedia sentences (≥ 60 characters) with suite contexts (HoVer 401, ANLI 154, BRIGHT 7, ToolRet 1); those rows carry uniform [0.5, 0.5] placeholder labels, so they hold no answer information.

runs/jev-27b

Same files, engine, suite, mechanical translation, capacity and training data as runs/jev-9b (the model is the same recipe on Qwen/Qwen3.8-27B @ 1d4bf0f2: rank-16 LoRA of 108,789,760 parameters and a 122,904-parameter head; 25,624,722,968 served parameters; temperatures noul 1.0143, choice 1.0161). On the model's own held-out test rows the engine matches our reference inference to max |Δp| 5.1e-07 with identical argmax (300 of 300).

area (skill)	Knowledge & Reasoning	Language	Retrieval & Classification	Tools	Arts
JEV-27B	0.398	0.574	0.537	0.735	0.399

Two phases, same output file: rows 1–2,361 (BFCL and the start of ToolRet) with the kit's sequential runner (environment-sequential.json; payload_sha256 of every finished row checked against the imported suite before resuming, 0 mismatches), rows 2,362–150,759 with source/fast_run.py, 64 requests per batch (environment-batched.json). Agreement check: the 1,694 phase-1 BFCL requests re-run through the batched driver give 4,768 of 4,768 identical answers (max |Δp| 0.016). Phase-2 total_wall_ms is batch time / 64. Hardware as above (1 × NVIDIA B200).

runs/jev-gemma4-26b-a4b

autotrust/JEV-Gemma4-26B-A4B @ 603bfed4: google/gemma-4-26B-A4B-it @ 4d7ae498 (mixture of experts, ≈ 4 B active parameters per token) + LoRA adapter (merged in memory at load) + 24-slot fp32 head (67,608) soft-capped at 30 like Gemma's logits; read-out at the last token of <bos> + the bare-v1 template; temperatures noul 1.0027, choice 1.0172.

area (skill)	Knowledge & Reasoning	Language	Retrieval & Classification	Tools	Arts
JEV-Gemma4-26B-A4B	0.430	0.636	0.679	0.697	0.415

Execution. One phase: all 150,759 rows with source/fast_run.py (64 requests per batch, shared length-sorted forward passes, right padding without an attention mask, exact for this causal model), environment-batched.json. source/jev_engine.py is the file that ran (sha256 9f01ca19…); the kit submission carries a later version that only adds an option for unmerged-adapter bundles (not used by this model). Phase total_wall_ms is batch time / 64. Hardware 1 × NVIDIA B200, torch 2.13 + CUDA 13.0, transformers 5.16.1, peft 0.21.0, kit 87d4650.

Training data. TypeSafe Jev 1.13 teacher distributions (the public jev-distill-corpus-v3) and ground-truth decision data from public datasets. The ground-truth data includes the public training splits of 12 datasets whose test splits the suite uses: BANKING77, CLINC150+OOS, WinoGrande, HellaSwag, GSM8K, ContractNLI, Amazon ESCI, HoVer, RAGTruth, When2Call, New Yorker caption matching and Habermas Machine. No test split of any suite benchmark was used; suite items were excluded before training; checkpoint selection and temperature fitting never used suite data. training-overlap-check.json: no whole suite state appears in the training data; the remaining instruction and sentence matches are the fixed per-benchmark templates.

Not affiliated with TypeSafe AI.