Zero-Token LLM Decision Engine
Microsecond Decisions.
Zero Model Weights. 100% Invariant.
Why spend 300 ms and $0.02 waiting for an LLM to classify or route a request? Zev solves order bias, temperature calibration, and guardrail abstention in Rust with single-digit microsecond latency.
Presets:
•
Decision Outcome
Ready
Press Evaluate Decision to view microsecond zero-token classification.
Primary Selected Destination
billing
ok
Calibrated Confidence
94.8%
T=2.179 Scaling
Execution Speed
6.2 µs
~48,000x faster than LLM
Calibrated Probabilities (100% Invariant)
Latency Realities (Log Scale)
⚡ Zev SIMD:
5.8 µs (0.0058 ms)
Together Tev1-4B:
300,000 µs (300 ms)
GPT-4o / Claude Haiku:
450,000 µs (450 ms)
🛡️ 100% Order Invariance Verification Suite
Standard LLMs suffer from severe position bias (often preferring the first or last option, with flip rates up to 35% when options are reordered). Zev mathematically isolates option attention slots, guaranteeing an exact 0.0% Permutation Flip Rate.
Permutation Stress Engine
Test permutation orders side-by-side to verify logits never drift.
Click "Run Invariance Test" to fire 5 different option orderings against Zev.
🤝 Together AI Tev1 Drop-In Evaluator
Zev is a wire-compatible, zero-token replacement for Together AI's tev1-4B-experimental format.
Drop in your State, Question, and lettered Options.
Tev1 Response
Run Tev1 evaluation to see output.
Compatible with
curl /v1/tev1 & Together prompt parser.
💻 Quick Integration Snippets
Interact with Zev via cURL, Rust, Node.js / npx, or Python.
cURL (/v1/decisions)
curl -X POST http://127.0.0.1:8080/v1/decisions \
-H "Content-Type: application/json" \
-d '{
"state": "Production database is down with 500 connection refused errors",
"questions": {
"route": {
"type": "choice",
"instructions": "Route incident to responsible team",
"options": [
{"id": "billing", "description": "Payment invoices and billing"},
{"id": "infra", "description": "Database outages and 500 connection errors"},
{"id": "sales", "description": "Enterprise sales inquiries"}
]
}
}
}'
Rust (zev-rs crate)
use zev::ZevEngine;
let engine = ZevEngine::default();
let (destination, confidence) = engine.route(
"Credit card dispute and refund inquiry for invoice #9401",
[
("billing".into(), "Refunds and payment invoices".into()),
("tech".into(), "API errors and server issues".into()),
].into_iter().collect(),
)?;
println!("Destination: {destination} (Confidence: {confidence:.2})");
Zero-Install CLI (npx)
npx -y zev-rs route \
--state "Critical: production database is down with 500 connection refused errors" \
--routes '{"billing": "Invoice inquiries", "infra": "Database outage and cluster errors"}'