bkhubbard/zev-benchmarks

Zev: Zero-Token Decision Engine Benchmarks

The Zev Benchmarks dataset provides standardized, reproducible evaluation schemas for zero-token decision engines, fast intent routers, and schema classifiers.

It is designed to evaluate four fundamental properties where standard autoregressive LLMs (like Llama, Mistral, or GPT) frequently fail or add excessive cost:

1. Option Order Bias & Invariance: Permuting candidate options should not change the decision or logits.
2. Probability Calibration: Calibrated confidence matching true accuracy (minimizing Expected Calibration Error).
3. Strict Abstention Guardrails: Refusing out-of-domain or evidence-free inputs via __insufficient__ rather than hallucinating.
4. Microsecond Evaluation Latency: Measuring execution speed against standard 300 ms – 1.2 s LLM generation.

---

Dataset Structure

Each row is a structured JSON record:

{
  "id": "zev_curated_0001",
  "task": "intent_routing",
  "state": "Customer requested an immediate full refund for invoice INV-2024-904 because the desktop app crashed on start.",
  "question": "Route customer request to the correct department",
  "options": [
    {"id": "billing", "description": "Refunds, payment disputes, invoices, subscription cancellations"},
    {"id": "tech_support", "description": "Crash reports, API errors, software installation bugs"},
    {"id": "sales", "description": "Enterprise license agreements and new tier upgrades"},
    {"id": "security", "description": "Vulnerability disclosures and compromised credentials"}
  ],
  "ground_truth": "billing",
  "difficulty": "easy",
  "split": "test"
}

---

Benchmark Task Categories

Task	Description	Evaluation Objective
intent_routing	Multi-class routing across customer support, billing, enterprise sales, and security.	High precision classification with zero token latency.
guardrail_abstention	Adversarial, out-of-domain, or evidence-free contexts.	Engine must output __insufficient__ instead of guessing.
order_invariance_eval	Paired test cases with candidate options reversed and shuffled.	Permutation Flip Rate must be 0.0%.
medical_triage	Clinical differentials (e.g. McBurney appendicitis vs gastroenteritis vs nephrolithiasis).	Multi-symptom semantic isolation and negation handling.
incident_response	Production infrastructure alerts (Postgres replication, Cloudflare 521, Kubernetes OOM).	Accurate SRE triage under technical noise.
tev1_eval	Drop-in Together AI tev1-4B-experimental schemas.	Wire parity with Together AI prompt formatting.

---

How to Load in Python

Using the datasets library from Hugging Face:

from datasets import load_dataset

dataset = load_dataset("json", data_files={
    "train": "train.jsonl",
    "test": "test.jsonl"
})

print(dataset["test"][0])

---

Comparative Results: Zev vs. Typical LLM API

Metric	Zev (zev-rs)	Together tev1-4B	Llama-3-8B
Latency	5.8 µs	300,000 µs (300 ms)	450,000 µs (450 ms)
Model Weights	0 MB	8,000 MB	16,000 MB
Permutation Flip Rate	0.0%	8.4%	14.2%
Cost per 1M Decisions	$0.00	$20.00	$30.00
Abstention Support	Native __insufficient__	Prompt dependent	Prompt dependent

---

Citation & Repository

@software{hubbard2026zev,
  author = {Brandon Hubbard and zev-rs contributors},
  title = {Zev: High-performance, 100% order-invariant zero-token LLM decision engine},
  url = {https://github.com/bhubbard/zev-rs},
  year = {2026}
}