/large-decision-models ● Active Homelab Node · Zero Public Ports

Large Decision Models (LDMs) & Homelab Neural Classification Engine

System 1 Non-Generative Intelligence · Live 6-Primitive Benchmark · 2026 LDM Landscape Matrix · Self-Hosted Homelab Architecture

1. What is a Large Decision Model (LDM)? System 1 vs System 2

SYSTEM 1 Large Decision Models (LDMs)

Non-generative encoder architectures specialized strictly in classification, multi-class triage, continuous scoring, and policy enforcement without autoregressive text generation.

⚡ Latency: ~25ms – 40ms warm / ~200ms cold 🎯 Output: Deterministic classes, bounds, relevance vectors 🛡️ Safety: Zero autoregressive hallucination

SYSTEM 2 Generative Frontier LLMs

Autoregressive foundation models designed for open-ended prose, multi-step synthetic reasoning, and complex code generation, producing one token at a time.

⏳ Latency: ~1,200ms – 2,500ms ✍️ Output: Unconstrained prose, code, reasoning narratives 💰 Cost: $2.00 – $50.00 / 1M tokens

2. Live Homelab Benchmark (Dave's ModernBERT Classifier)

Interactive 2-column command workbench executing live neural classification through Dave's on-premise container. Displays real measured round-trip latency.

● Latency: -- ms

// 01. INPUT TEXT PAYLOAD

Ready for query

TELEMETRY // DECISION BENCH

WAITING

Click 'Execute Homelab Inference' above to evaluate payload through Dave's Local ModernBERT classifier...

Click "Execute Homelab Inference" to dispatch payload to Dave's ModernBERT classifier...

3. The 2026 Large Decision Model (LDM) Landscape

Objective comparative breakdown of local self-hosted open-source models versus commercial managed decision APIs.

Market Matrix

Class / Tier	Model / Engine	Deployment Runtime	Typical Latency	Cost Model	Primary Production Use Case
LOCAL HOMELAB	Dave Laya Engine ModernBERT-large (421M)	Self-Hosted Linux Node	~25ms – 40ms	$0 / mo	Sub-second triage, PII tripwires, zero-egress routing
LOCAL OPEN-SOURCE	BGE-Reranker-Large BAAI / Hugging Face	Local PyTorch / ONNX	~40ms – 65ms	$0 / mo	RAG semantic passage reranking, context pruning
LOCAL OPEN-SOURCE	DeBERTa-v3-large-NLI Microsoft Research	Local Container / CPU	~50ms – 80ms	$0 / mo	Zero-shot NLI entailment, premise classification
MANAGED CLOUD API	OpenAI Decisions API OpenAI Cloud	Cloud REST Endpoint	~140ms – 240ms	$0.05 – $0.20 / 1k	Enterprise policy checks, structured output gating
MANAGED CLOUD API	Cohere Rerank 3.5 Cohere AI	Cloud REST API	~120ms – 200ms	$2.00 / 1k queries	Enterprise multi-lingual document & search reranking
ENTERPRISE VPC / API	TypeSafe Jev TypeSafe AI	Managed VPC / Cloud	~70ms – 130ms	Enterprise Tier	Deterministic type schema enforcement & routing

4. The Homelab Architecture Story & Private Decision API

How this physical homelab node operates securely and evolves into an enterprise-ready private Decision API.

● Homelab Infrastructure

01 // HARDWARE & TOPOLOGY

Physical Hardware & Internal Loopback

Hosted on a dedicated bare-metal Linux server ('Dave') running Ubuntu 26.04 LTS with an NVMe system drive and 14 GiB RAM. The ModernBERT 421M parameter model runs in a self-contained Docker container, processing requests over internal loopback at 192.168.50.152:8095.

02 // ZERO-TRUST DEFENSE

Zero-Trust Mesh & Tunnel Ingress

The host has zero public open ports and sits behind consumer double-NAT. Management and administration route exclusively through an encrypted Tailscale WireGuard mesh (100.81.47.31). Public ingress is provided strictly via an outbound Cloudflare Zero Trust tunnel with edge DDoS/WAF protection.

03 // EVOLVING TO AN API

Transforming into an Authenticated Enterprise API

By pairing DaveRouter's token-authenticated proxy layer with the ModernBERT container, this homelab setup easily functions as an authenticated REST API for internal chatbots, automated triage pipelines, and enterprise tools at $0/mo marginal cloud cost.