The Herald pattern is most valuable where speed, reliability, auditability, or cost constraints make the conventional LLM-as-brain approach impractical or unacceptable. With 46k+ lines of deterministic code and ten specialist model seats, the architecture is production-scale.
Consumer
Voice Assistants
Local, private, fast. No cloud API dependency. Ten specialist model seats on a single consumer GPU via sequential relay. <100ms deterministic routing through the 5-stage cascade. The LLM adds polish, not intelligence.
Enterprise
Enterprise Chatbots
Auditable decision paths for compliance teams. Every routing decision is traceable code, not black-box reasoning. SOC 2 and HIPAA-friendly by architecture.
Regulated
Safety-Critical Systems
Medical triage, legal document analysis, financial advisory. Where LLM hallucination isn't an inconvenience -- it's a liability. Deterministic routing eliminates hallucinated tool calls entirely.
Hardware
Edge & IoT Deployment
A 1B parameter renderer handles most turns. Sequential relay loads specialist seats on-demand to conserve VRAM. Most turns skip inference entirely. Runs on devices where you can't afford cloud latency, bandwidth costs, or API rate limits.
Infrastructure
Multi-Agent Orchestration
Ten specialist model seats coordinate through a shared world-state model, not competing LLM context windows. BG1 streaming delivers real-time progress at 15%, 60%, 95%. Deterministic routing across 171 modules means code orchestrates, not prompt engineering.
Research
Self-Improving Systems
The CRSIS path built on Concept D. A deterministic brain with a world-state model can version, diff, test, and roll back its own decision logic. Self-modification that is safe because it is transparent.
Trust
Evidence-Critical AI
Every claim traced to evidence. Medical, legal, financial — where 'the model said so' is not acceptable. Skeptic tags every output as observed, recalled, inferred, or guessed.