Intelligence Is a Component. Control Is an Architecture.
Mythos Systems uses commercial, open-weight, local, hosted, and browser-local models according to task, privacy, cost, and hardware.
Models may retrieve, summarize, compare, classify, plan, and coordinate. Deterministic systems validate, authorize, execute, observe, and preserve authoritative state.
Model Classes
Hosted Commercial Models
High capability and broad reasoning when data policy permits and result quality matters most.
Open-Weight Models
Control, customization, local deployment, privacy, and cost predictability.
Local Desktop Models
Private inference on user-controlled hardware without network dependency.
Browser-Local Models
WebLLM and WebGPU for optional client-side inference with explicit user consent.
Model Routing
Model type is selected according to task difficulty, latency, privacy, cost, hardware, context size, tool use, evaluation results, availability, and safety requirements.
RAG and Knowledge
Retrieval pipelines include document ingestion, chunking, embeddings, semantic search, reranking, citations, and knowledge boundaries. Public assistants use approved corpus only.
Agents and Tools
Agentic workflows include multi-step planning, iterative retrieval, tool use, verification, budgets, and stopping conditions. Tools require explicit permissions and deterministic validation before consequential actions.
Memory and Evaluation
Memory systems define conversation state, task memory, project knowledge, provenance, and deletion. Evaluation covers factuality, grounding, tool selection, refusal behavior, prompt injection, regression, cost, and latency.
Security and Prompt Injection
Models do not silently become authoritative state. Deterministic systems remain responsible for permissions, policy, validation, execution boundaries, destructive actions, transactions, rollback, and audit.