Prometheus Architecture

Simplified map: one main loop, one validation layer, one runtime layer. Details are grouped instead of drawn as crossing wires.

Simplified Prometheus architecture map A simplified dark SVG diagram showing the Prometheus research loop, validation lanes, local model services, GPU acceleration, databases, and Hermes fork dependencies. MAIN RESEARCH LOOP VALIDATION + EPISTEMIC GATES LOCAL RUNTIME + GPU UTILIZATION refill dispatch write result apply follow-ups: transfer questions, mutations, synthesis, compression challenge + calibrate models + GPU 1. Questions prometheus.db curiosities scores + lane priorities parent lineage evidence_depth what should run next? 2. Task Queue kanban.db task_refiller 2m priority 4/3/2/1/0 RAG prior feed blind-lane stamps turn questions into cards 3. Workers stateless LLM sessions ~20 concurrent real code + experiments mimo · A1 · deepseek · free confidence cap 0.85 workers die; DB memory stays 4. Intake worker_results write_worker_result result_bridge 10m quality gate ≥ 40 artifacts + manifest raw findings land here 5. Knowledge experiments + claims maturity tiers mechanisms transfer_tracking claim_evidence what the system learned External Truth BENCH3 + calibration known-answer probes · domain accuracy validates answers Claim Gauntlet retests · attacks · arbitration SA · circularity · method/code challenges evidence before belief Novelty + World literature · discovery shelf finder · scholar_search · world_grounding separates novelty from robustness Synthesis compression + topology auto_tune · outcome routing turns structure into questions Agents-A1 :8001 local vLLM Qwen3.5-MoE transfer/exp lanes + attacks GPU Embed :9150 Qwen3-Embedding-0.6B RAG · compression · similarity GPU Sklearn transparent .pth hook LR · PCA · Scaler · KNN vllm-env + gpu_run torch/CUDA runtime load-bearing; workers depend on it Hermes Fork 15-file patch stack routing + guard + venv fixes Legend workers/services/scripts databases/artifacts/indexes queue/generation bus epistemic gates/fork risk external/local GPU runtime solid = main flow dashed = grouped support/validation, not every call

Main loop

  • • Questions live in `prometheus.db`; runnable cards live in `kanban.db`.
  • • Workers are stateless; the databases are the memory.
  • • Intake turns worker results into experiments, claims, mechanisms, and follow-up questions.

Trust layer

  • • BENCH3 checks answer correctness; transfer survival checks mechanism portability.
  • • Retests, attacks, arbitration, clean-room, circularity, and method-code checks challenge claims.
  • • Literature and world-grounding keep novelty separate from robustness.

Utilization

  • • Agents-A1 handles local reasoning lanes and some attacks.
  • • GPU Embed powers RAG, synthesis, and similarity.
  • • GPU sklearn and `gpu_run` give workers transparent GPU-backed compute.

Prometheus architecture diagram · Simplified from ~/.hermes/docs/architecture-map.md · Offline HTML/SVG