ZorQelis Decision Suite (ZDS-1)

One benchmark for decision models. A decision model reads a situation (the state), a question and a fixed list of options, and returns one option with a calibrated confidence. No free text. ZDS-1 measures that job on 22,864 cases in three tracks, and scores every system with the same harness, the same metrics and the same pricing rule.

Cases	22,864 in 3 tracks
Metrics	accuracy (Wilson 95% CI), ECE, AURC, p50 / p95 latency, cost per 1M decisions
Score	ZDS-1 = mean of the three track accuracies, x100
Systems scored	VerQen, Laya, Jev 1.13
Released	27 September 2026, by ZorQelis AI

Leaderboard

System	ZDS-1	Jev-parity	Laya-parity	AGY-32	p50 latency	$ / 1M decisions
typesafe/jev-1.13	78.4	60.6	77.9	96.9	409 to 443 ms (API)	15 to 29 (billed)
zorqelis-ai/VerQen	73.9	93.7	75.9	52.0	4.7 to 7.3 ms	1.7 to 4.0
convaiinnovations/laya	52.6	19.3	63.1	75.5	26 to 58 ms	6 to 13

Calibration (expected calibration error, lower is better): VerQen 0.095 / 0.039 / 0.143, Laya 0.521 / 0.124 / 0.052, Jev 0.133 / 0.091 / 0.013 on Jev-parity / Laya-parity / AGY-32.

No system wins every track. Jev leads the suite on open-domain judgement. VerQen leads by a wide margin when a decision is governed by rules written in the input, and is 60 to 90 times faster. Laya sits between them on AGY-32. The tables below show where each one is strong.

The three tracks

Track	Cases	What it asks	Where the cases come from
Jev-parity	7,840	Apply the policy written in the state to the case: arithmetic, dates, counting, counterfactuals, distractors, prompt injection, missing and conflicting information, long inputs	Generated by ZorQelis AI's decision engine on held-out phrasings, 22 business domains
Laya-parity	14,000	Real classification and reading tasks, 2,000 each	Test splits of ag_news, banking77, BoolQ, emotion, MASSIVE (en), SST-5, XNLI (en)
AGY-32	1,024	Open-domain judgement with no written rules: 32 domains x 32 hand-written cases	Authored for this suite, outside every training generator used by VerQen

Jev-parity, by capability (accuracy %)

Capability	VerQen	Laya	Jev 1.13
Arithmetic	88.2	21.8	64.3
Conflicting information	100.0	0.6	7.3
Counterfactual	91.9	34.3	87.7
Counting	94.8	24.8	76.2
Dates	89.3	23.0	75.8
Distractors	96.1	18.1	83.5
Grounded rules	91.8	29.0	82.6
Prompt injection	93.0	16.2	53.2
Long state	86.8	26.5	83.1
Missing information	99.6	1.9	0.7

Missing and conflicting information expect the "gather more information" action. Jev and Laya almost never take it; they commit to a decision.

Laya-parity, by dataset (accuracy %)

Dataset	VerQen	Laya	Jev 1.13
ag_news	89.8	93.2	89.6
banking77 (77 intents)	85.5	35.8	79.5
BoolQ	71.2	78.0	91.8
emotion	91.5	58.2	59.5
MASSIVE en (59 intents)	77.6	43.5	80.1
SST-5	47.4	47.0	57.2
XNLI en	68.5	86.2	87.5

AGY-32, by domain (accuracy %, 32 cases each)

Domain	VerQen	Laya	Jev 1.13
agent_tool_routing	65.6	78.1	100.0
b2b_sales_lead_scoring	50.0	68.8	100.0
chemistry_reaction_classes	53.1	84.4	93.8
code_bug_severity	46.9	65.6	90.6
commonsense_causal_reasoning	56.2	78.1	96.9
content_moderation_toxicity	84.4	100.0	100.0
customer_support_triage	34.4	93.8	100.0
cybersecurity_phishing	68.8	75.0	100.0
data_schema_type_inference	37.5	53.1	100.0
devops_system_monitoring	56.2	84.4	100.0
ecommerce_product_category	65.6	71.9	100.0
education_student_mastery	28.1	93.8	100.0
email_urgency_scoring	28.1	62.5	78.1
financial_fraud_detection	43.8	71.9	100.0
formal_boolean_logic	53.1	62.5	100.0
game_chess_tactics	37.5	50.0	96.9
geography_navigation_routes	28.1	43.8	96.9
hotel_hospitality_concierge	46.9	75.0	100.0
hr_candidate_screening	84.4	100.0	100.0
insurance_claims_processing	46.9	65.6	93.8
legal_contract_clause	75.0	96.9	100.0
llm_jailbreak_safety	68.8	93.8	100.0
logistics_freight_shipping	40.6	65.6	100.0
math_arithmetic_reasoning	28.1	34.4	100.0
medical_triage_acuity	28.1	62.5	65.6
natural_language_inference	71.9	100.0	100.0
physics_mechanics_forces	40.6	81.2	100.0
rag_passage_relevance	43.8	96.9	100.0
sentiment_analysis	87.5	96.9	100.0
smart_home_iot_automation	56.2	84.4	96.9
social_conversation_intent	56.2	87.5	100.0
temporal_event_sequencing	53.1	37.5	90.6

Run it

git clone https://huggingface.co/datasets/zorqelis-ai/zds-1 && cd zds-1
pip install git+https://github.com/ZorQelis-AI/verqen    # the harness and the VerQen adapter
pip install datasets huggingface_hub

# a local model: VerQen straight from the Hub
python harness/zds1.py --adapter verqen --checkpoint zorqelis-ai/VerQen --out-dir results/verqen

# Jev through OpenRouter's Decisions API
OPENROUTER_API_KEY=... python harness/zds1.py --adapter jev_openrouter --model typesafe/jev-1.13 --out-dir results/jev

The first run rebuilds the full Laya-parity track (see Licensing) and refuses to score if the rebuilt file does not match data/SHA256SUMS.json. harness/compare.py A.json B.json gives a paired comparison of two result files: the difference with a 95% bootstrap CI and an exact McNemar test per capability.

Every case is one JSON line:

{"id": "banking77/355", "track": "laya_parity", "capability": "banking77", "domain": "laya_bench", "state": "What is the auto top up function and where can I find that at?", "question": "What does the customer want to do?", "options": ["activate my card", "age limit", "..."], "answer": "automatic top up"}

How the scores were produced

• Same inputs for everyone. Each system receives the identical state, question and option list. The harness sends one decision per call and records the system's own confidence.
• Metrics. Accuracy with a Wilson 95% interval; top-label expected calibration error (10 bins); AURC, the area under the risk-coverage curve, which says whether a confidence is worth routing on.
• Hardware. VerQen and Laya ran on the same NVIDIA RTX 5060 Ti (16 GB), one decision at a time. Jev ran through OpenRouter's Decisions API (POST /api/alpha/decisions), served snapshot typesafe/jev-1.13-20260917, 16 requests in parallel, 0 failed calls. Jev's latency includes the network round trip from India, so it overstates the model's own compute time.
• Cost. Local models are priced at a cloud L4-class GPU, $0.80 per hour, at their measured latency. Jev's cost is what OpenRouter billed, per call: $0.64 for all 22,864 cases.
• Versions. VerQen zorqelis-ai/VerQen (232M parameters); Laya convaiinnovations/laya through the laya package 0.3.20; Jev 1.13. Runs dated 25 to 27 September 2026.

Read the results with these facts in mind

• Jev-parity is home ground for VerQen. ZorQelis AI built it, with the same decision engine that generated part of VerQen's training data. The test uses held-out phrasings and domains, but the task format is VerQen's own. Two of VerQen's fixes (conflicting and missing information) were chosen after seeing earlier Jev-parity scores.
• VerQen saw the train splits of six of the seven Laya-parity datasets (all but XNLI; MNLI, XNLI's English training set, is in its data). It never saw a test split, and training rows that exactly matched a case were removed. VerQen's pretraining corpus was 13-gram filtered against evaluation sets, but emotion, banking77, SST-5 and MASSIVE joined the evaluation after that filter ran. Laya's and Jev's training data are not published.
• AGY-32 is the least favourable track for VerQen, and it is included for that reason: it asks for judgement from general knowledge with no rules in the input.
• The question wording is ours for every track. Laya's and Jev's own published numbers used their own prompts and will not reproduce exactly.

Licensing

Part	Licence
Jev-parity, AGY-32, the harness	CC BY 4.0 (cases), Apache-2.0 (code). Jev-parity's long-state filler text is from FineWeb-Edu (ODC-By).
Laya-parity: banking77, MASSIVE	CC BY 4.0, from their original authors
Laya-parity: BoolQ	CC BY-SA 3.0, from its original authors
Laya-parity: ag_news, emotion, SST-5, XNLI	Not redistributed (unknown or non-commercial licences). harness/build_laya_parity.py downloads them from their original hubs and rebuilds the exact cases, checked by SHA-256.

Citation

@misc{zds1_2026,
  title  = {ZorQelis Decision Suite (ZDS-1): a benchmark for decision models},
  author = {{ZorQelis AI}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/zorqelis-ai/zds-1}
}