DCA-01 - Deterministic Cluster Architecture
From monolithic model to deterministic AI ecosystem
Template: Just-in-Sequence Manufacturing (Automotive) + Triple-Modular-Redundancy TMR (Avionics) + CAN-Bus / Blackboard + LRU Principle
Version: 1.2 COMPLETE DOSSIER | Author: Emanuel Schaaf | Status: Implementable
What is this?
DCA-01 is a build-ready specification for an AI infrastructure that is:
- Self-optimizable: Each module learns from its own failures via local LoRA, no global retraining
- Easier to train: Local training, isolated weights, no catastrophic forgetting
- Faster & more accurate: Only relevant experts active, parallel inference, calibrated confidence c_i
- Auditable: Every path logged with dag_id, model_hash, provenance - EU AI Act ready
Instead of one giant model carrying all world knowledge for every token, a cluster of physically separated, domain-specialized base models is orchestrated via a deterministic data bus.
Core principle: Knowledge transfer happens not through mixed weights, but through typed data exchange on a Shared Context Board (Blackboard). Like CAN-Bus in a car.
Architecture
[User Request] -> [TMR Router 2-out-of-3] -> [Task Decomposer DAG] -> [Semantic Bus / Blackboard] <-> [Experts: Physics/HPC, Code/Mid-GPU, Language/CPU] -> [Fusion Layer Delta Check + Micro-Iterations] -> [Final Answer + Audit Log + Provenance] -> [RL Self-Optimization LoRA]
See docs/diagrams/ for DE/EN architecture images.
Repo Structure
/schemas/ - All JSON Schemas (Router, DAG Node, Blackboard Pattern, Expert Output E_i, Final Answer)
/src/router/ - TMR Router A,B,C + Voting
/src/decomposer/ - DAG Builder
/src/blackboard/ - Semantic Bus / Blackboard Publish/Subscribe
/src/experts/ - Isolated domain models: physics, code, language
/src/fusion/ - Deterministic Fusion with Delta Check
/src/observability/- Prometheus metrics, OpenTelemetry tracing
/configs/ - Hardware mapping, budgets
/docs/ - Full dossiers DE/EN + PDFs
Quick Start MVP
# Phase 0 MVP - 1 Router + 2 Experts + In-Memory Blackboard
docker-compose up --build
python src/router/tmr_router.py --request "Write Python simulation"
python -m pytest tests/
Formal Output
Every expert returns: E_i(T_i) = (y_i, c_i, t_i, model_hash, provenance)
Fusion checks: delta = |c_physics - c_code|, type_check, unit_check. On fail: Constraint -> Re-run only failing module, max 3 iterations.
Dossiers
License
MIT - Open Spec for industrial implementation
Contact:
- 📧 Contact