FMS — Foundation Model Service for CoreSense

A reference implementation of a CoreSense cognitive structure for deploying AI Foundation Models (FM)—both open-source models for local deployments and closed systems accessed through commercial APIs—as typed cognitive organs inside a CoreSense awareness pipeline.

---

Overview

Modern AI Foundation Models (LLMs, VLMs, and multimodal FMs) are powerful but structurally incomplete as autonomous reasoners: they lack grounded consistency guarantees, cannot self-verify commutativity against the world, and are prone to hallucination.
CoreSense treats an FM not as an "intelligent system" but as a typed cognitive organ—a specialised component placed at well-defined stages of the awareness pipeline, with its outputs entering the knowledge base only through consistency-maintaining integration.

This repository provides:

Artefact	Purpose
README.md	Top-level orientation and catalog index
docs/FMS_Architecture.md	Full architecture description, SysMLv2 models and diagrams
docs/catalog/	One document per FM use pattern

---

The CoreSense Awareness Pipeline

 ┌─────────┐ ┌─────────┐ ┌───────────────┐ ┌───────────┐ ┌──────────┐
│ Sense │───▶│ Extract │───▶│ Model-Propose │───▶│ Integrate │───▶│ Evaluate │
└─────────┘ └────┬────┘ └──────┬────────┘ └─────┬─────┘ └────┬─────┘
│ │ │ │
◀─ FM eligible ─▶ ◀─ FM eligible ─▶ ◀── FM excluded ──▶

Foundation Models are admitted only at the Extract and Model-Propose stages.
They are deliberately excluded from Integrate and Evaluate, because a corpus-grounded model cannot check its own commutativity against the world.
Every FM exertion yields a provenance-tagged candidate assertion plus a process record, entering the knowledge base only through the consistency-maintaining integrator.

---

Catalog of FM Use Patterns

Eight canonical patterns are defined, spanning both deployment mode and epistemic role:

#	Pattern	Epistemic Role	Deployment	Document
1	Abductive Engine	High-recall candidate model fragment proposer	API / Local	P1
2	Feature Extractor	Signal → symbolic feature vector	Local / Edge	P2
3	Knowledge Elicitor	Unstructured text → structured KB fragments	API	P3
4	NL Interface	Natural-language ↔ formal query bridge	API / Local	P4
5	Code/Logic Generator	Produce executable symbolic programs	API	P5
6	Similarity Oracle	Semantic similarity and analogical retrieval	Local	P6
7	Hypothesis Generator	Scientific/causal hypothesis draft	API	P7
8	Explanation Narrator	Translate formal derivations into NL explanations	API / Local	P8

---

Supported FM Deployment Modes

Mode	Examples	Key Characteristic
Local OSS	Llama 3, Mistral, Phi-3, Qwen 2, DeepSeek	Full data residency, tunable latency
Commercial API	Claude (Anthropic), GPT-4o (OpenAI), Gemini (Google), Kimi K3 (Moonshot)	State-of-the-art capability, usage-metered
Cloud ML Platform	Azure ML, AWS Bedrock, Google Vertex AI	Enterprise governance, managed endpoints
Hybrid (draft+verify)	Local draft → Cloud verify	Cost-latency optimisation

---

Key Design Principles

1. Type discipline — every FM is assigned a CognitiveOrganType with an explicit epistemic contract.
2. Provenance — every FM output carries a ProvenanceRecord (model id, version, temperature, prompt hash, timestamp).
3. Quarantine — FM outputs enter a CandidateBuffer; only after passing the ConsistencyIntegrator do they enter the KnowledgeBase.
4. Pipeline placement — FM organs are wired via FMPort connectors only to Extract and ModelPropose pipeline blocks.
5. Observability — each exertion emits a ProcessRecord for audit.

---

Repository Structure

fms/
├── README.md ← this file
├── docs/
│ ├── FMS_Architecture.md ← full architecture + SysMLv2 + diagrams
│ └── catalog/
│ ├── P1_AbductiveEngine.md
│ ├── P2_FeatureExtractor.md
│ ├── P3_KnowledgeElicitor.md
│ ├── P4_NLInterface.md
│ ├── P5_CodeLogicGenerator.md
│ ├── P6_SimilarityOracle.md
│ ├── P7_HypothesisGenerator.md
│ └── P8_ExplanationNarrator.md
└── LICENSE

---

References

• CoreSense project: https://github.com/CoreSenseEU
• Kimi K3 (MoonshotAI): https://www.moonshot.cn
• Azure ML Foundation Models: https://learn.microsoft.com/en-us/azure/machine-learning/how-to-use-foundation-models?view=azureml-api-2
• SysMLv2 specification: https://www.omg.org/spec/SysML/2.0
• Ollama local model serving: https://ollama.com