Evaluating formal ontologies in < 1ms with proven subsumption monotonicity and zero hallucination. At $0.042 per 1M input tokens, outputs are always 100% free.
Generative LLMs sample tokens autoregressively one by one with unconstrained softmax pools. Leio evaluates formal knowledge graphs in a single forward pass with axiomatic exclusion.
Head-to-Head Latency & Reliability Race
Live simulated benchmark against cloud-hosted autoregressive models
Legacy: Generative LLM
1,450 ms
Output Paradigm:Autoregressive tokens (1 by 1)
Hallucination Risk:8–14% (Stochastic sampling)
Math Guarantee:None (Unconstrained softmax)
Cost per 1M Decisions:$15,000.00 (Input + Output)
Client Offline:Impossible (Cloud GPUs only)
Idle — click "Run Live Latency Race"
⚡ Leio: Ontology-Grounded System 1
< 0.001 ms (0.8 μs)
Output Paradigm:Single forward pass over formal DAG
Hallucination Risk:0.000% (Axiom verified)
Math Guarantee:P(child) ≤ P(parent) everywhere
Cost per 1M Decisions:$0.042 (Outputs 100% Free)
Client Offline:189 KB WebAssembly in Browser
Ready in RAM (>500k decisions/sec)
REAL-TIME INDUSTRIAL AUTONOMY
Zero-latency dispatch for autonomous physical systems.
Watch a live System 1 agent evaluate high-speed 200kHz barcode sensor streams and trigger pneumatic diverter gates across supply chain chutes using formal GS1 axioms in <1ms.
🤖 Autonomous Cross-Dock Sorter Agent● Live System 1 Agent
Dispatching pallets across high-speed conveyor chutes using GS1 ontology axioms.
Autonomous Agent Audit Log (Zero-Hallucination Loop)0 ms
[System Ready] Leio decision engine initialized with GS1 Web Vocabulary DAG (63 classes).
[Agent Idle] Press "Run Live Mission" to watch autonomous barcode intake and conveyor diverts.
STRUCTURED DECISION PRIMITIVES
Five primitives. Zero prose.
Instead of prompting an unconstrained language model with paragraphs of instructions, you formulate decisions using five mathematically rigorous primitives.
from olaya import OLaya
olaya = OLaya()
state = {
"gtin": "0614141000036",
"scan_text": "Grade A pasteurized whole milk, 1 gallon jug with tamper-evident seal. Keep refrigerated at 34-38 F.",
"sensor_temp_c": 3.2,
"carrier": "ColdExpress Logistics",
}
questions = {
"taxonomy_grounding": { "type": "concept", "root": "Product", "instructions": "Ground item into GS1 category." },
"requires_cold_chain": { "type": "subsumes", "concept": "FoodBeverageTobaccoProduct", "instructions": "Perishable handling rules?" },
"packaging_triage": { "type": "choice", "criteria": {"pallet_return": "reusable wooden pallet", "corrugated_bale": "cardboard carton"} },
"dock_urgency": { "type": "score", "criteria": ["ambient_low", "standard_dock", "expedited_perishable_cold_chain"] },
"seal_breach_alert": { "type": "noul", "instructions": "Any indication of seal damage or temperature excursion?" }
}
result = olaya.system_one(state, questions)
1. concept
Leaf taxonomy grounding with guaranteed monotonicity: P(child) ≤ P(parent) everywhere. Returns full ancestral lineage with no broken category math.
P(Product) ≥ P(FoodBeverage) ≥ P(Milk)
2. subsumes
Tests whether the item belongs to the principal down-set ↓C of a target category. Enables instantaneous PLC diverter triggers without querying external clouds.
↓C = {x ∈ Ο | x ⊑ C}
3. choice
Evaluates mutually exclusive options with category-theoretic exclusion. If two classes are disjoint, their joint probability is strictly zero.
A ⊓ B = ⊥ &implies; P(A ∩ B) = 0
4. score & noul
Ordered ordinal calibration for priority queues, alongside binary out-of-domain anomaly gating to catch sensor excursions and unknown SKUs.
E[Y] = ∑ i · P(level_i)
VERIFIABLE HARDWARE BENCHMARKS
Measured on bare metal. Proven at scale.
Tested on Apple Silicon MPS and single-core CPU on operational GS1 supply chain workloads. The WASM build runs entirely in-browser with zero network round-trip.
Runtime / Architecture
Parameters
Footprint
Median Latency
P95 Latency
Throughput
Deployment Target
🦀 Rust WASM (olaya-wasm) LIVE ✓
0 (pure index)
189 KB .wasm binary
< 1 μs
< 2 μs
>500,000 qps
In-browser client · zero network, zero server
Rust Arrow Lexical (olaya-core)
0 (pure index)
Arrow IPC Buffer
4.8 μs
5.2 μs
~210,000 qps
Embedded PLC · Conveyor Chute Divert
Custom BERT Decision Model
15.37M
4 layers, 256d
12.71 ms
14.52 ms
231.7 qps
Edge Scanner · Forklift Handheld Terminal
ModernBERT-base Backbone
149.91M
22 layers, 768d
199.10 ms
217.52 ms
14.9 qps
Control Tower · High-Fidelity Triage
RADICAL PRICING ECONOMICS
$0.042 per million input tokens. Outputs free.
Generative LLMs bill you for slow, verbose output tokens. Leio meters inputs only at Jev-compatible rates, making output decisions completely unmetered and free.