The AENEA ecosystem delivers high-precision semantic routing and System 1 decision engines built for the Human + Agent Economy. Route prompts across MoE backends, or execute complex enterprise logic with zero hallucinations and mathematical calibration.
74.8M System 1 Decision Engine
Orthogonal geometry for enterprise logic. Beats 421M dense models on semantic boundaries. Zero hallucinations, mathematically calibrated probabilities.
Hugging Face ↗74.8M Parameter Semantic Router
24-layer Cittern-2.0 architecture. Classifies prompts into 9 high-precision domain tokens in a single forward pass for dynamic MoE dispatch.
Hugging Face ↗50M Parameter Edge Router
Production-grade 50M router. Ideal for ultra-low-latency edge dispatch where every millisecond counts. Still supported and maintained.
Hugging Face ↗Pinta-2.0 classifies every prompt into exactly one of nine reserved domain tokens — read from the vocabulary logits in a single forward pass. No calibration, no heuristics, no ambiguity.
"Review this contract for GDPR Article 17 compliance"<|reserved_30|> (Compliance) · Confidence: 0.993
Evaluated on 1,060 prompts across 9 domains (zero train overlap, English-only). Pinta-2.0 outperforms the previous generation across every metric while maintaining sub-10ms GPU inference latency.
| Domain | Precision | Recall | F1-Score |
|---|---|---|---|
| SyntaxDevOps | 0.94 | 0.76 | 0.84 |
| Code | 0.83 | 0.83 | 0.83 |
| Creative | 0.94 | 0.88 | 0.91 |
| RAG | 0.96 | 0.99 | 0.98 |
| Architecture | 0.93 | 0.94 | 0.93 |
| Math | 0.91 | 1.00 | 0.95 |
| Knowledge | 0.84 | 0.96 | 0.89 |
| Compliance | 0.99 | 0.97 | 0.98 |
| Chat | 0.99 | 0.97 | 0.98 |
| Weighted Average | — | — | 0.92 |
Evaluated on 250 specialist enterprise decision tasks. BHR-0.1's orthogonal manifold preserves semantic boundaries that dense models collapse, achieving superior contradiction detection at a fraction of the parameter count.
| Metric | BHR-0.1 (74.8M) | Laya (421M) | Advantage |
|---|---|---|---|
| Contradiction Detection | 98.0% | 82.2% | +15.8% (BHR Wins) |
| JSON Procedural Logic | 100.0% | 100.0% | TIE |
| Avg Latency (CPU) | 187.9 ms | 321.9 ms | 1.7x Faster |
| Calibration (Brier) ↓ | 0.0017 | 0.081 | Mathematically Honest |
| Overall Accuracy | 87.2% | 90.0% | 5.6x fewer params |
Production-grade routing and decision making with zero inference-time calibration. Pinta-2.0 ships with PyTorch, fp16, and ONNX artifacts. BHR-0.1 ships fully containerized.
import torch
from transformers import AutoTokenizer
from cartan_olm import CartanConfig, CartanLM
# Load artifacts
tokenizer = AutoTokenizer.from_pretrained("tokenizers/QT-Cittern-12k")
checkpoint = torch.load("pinta-2.0.pt", map_location="cuda")
config = CartanConfig.from_dict(checkpoint["config"])
model = CartanLM(config).to("cuda").eval()
model.load_state_dict(checkpoint["model_state_dict"])
DOMAINS = [f"<|reserved_{i}|>" for i in range(23, 32)]
domain_ids = [tokenizer.convert_tokens_to_ids(d) for d in DOMAINS]
def route(prompt: str) -> dict:
formatted = f"User: {prompt}\nAssistant: "
ids = tokenizer.encode(formatted, return_tensors="pt",
truncation=True, max_length=2048).to("cuda")
with torch.no_grad(), torch.autocast(device_type="cuda", dtype=torch.bfloat16):
logits = model(ids)["logits"][0, -1, :]
scores = logits[domain_ids]
probs = torch.softmax(scores, dim=-1)
best = torch.argmax(probs).item()
return {"token": DOMAINS[best], "confidence": probs[best].item()}docker run -p 8000:8000 aenea/bhr-0.1
Built on Block Householder Linear layers to constrain weights to the Stiefel manifold. This orthogonal geometry prevents representation collapse in small models, enabling state-of-the-art performance at a fraction of the parameter count.
CartanLM Architecture
Cittern-2.0 Architecture
Cittern-1.0 Architecture
This exact page is served byte-for-byte from both aenea.app and quartz.host — wherever you knock, the same ecosystem answers.