Version 0.1 - Early Draft for Discussion
The AI ecosystem is evolving rapidly, with countless tools, platforms, and frameworks emerging daily. Yet we lack a shared mental model—a common way to discuss where different capabilities sit, how they interact, and where the gaps might be.
This repository contains AILIS (AI Layer Interface Specification), a proposal exploring whether a layered model—similar to the OSI model in networking—might help us better understand and discuss AI system architectures.
In networking, the OSI model gave us a shared vocabulary. When someone says "Layer 3," everyone understands we're discussing routing and addressing, not physical cables or application logic.
Could something similar help in AI? We're not sure, but we think it's worth exploring together.
New to AILIS? Here's how to get oriented:
- Read the AILIS Primer (10 min) - Comprehensive overview of the 16-layer model
- Check the Cheat Sheet (2 min) - Quick reference for all layers
- Browse Case Studies - See how real systems map to the model
- Join the Discussion - Share your thoughts
Want to contribute?
- See CONTRIBUTING.md for the RFC process
- Check FEEDBACK.md for areas where we need input
- Review open issues for active discussions
Building the website locally?
# Install dependencies
pip install -r requirements.txt
# Preview site (with live reload)
mkdocs serve # http://localhost:8000We're proposing a 16-layer model (plus cross-cutting concerns) that attempts to map the AI stack from physical infrastructure up through application logic:
- L0 – Facilities & Power: Datacenters, power/cooling, physical security
- L1 – Compute Fabric: GPUs/TPUs/NPUs/CPUs, memory, interconnects
- L2 – System & Driver Runtime: CUDA/ROCm/Metal, device memory management
- L3 – ML Graph & Compilation: XLA/TVM/TensorRT-LLM/ONNX Runtime
- L4 – Numeric & Quantization: FP16/FP8/INT4, sparsity, calibration
- L5 – Tokenization & Encoders: BPE tokenizers, CLIP, audio patchifiers
- L6 – Model Parameters & Architecture: Base/foundation weights, MoE, diffusion
- L7 – Inference Engine & Decoding: Serving runtimes, caching, speculative decoding
- L8 – Context Construction & Prompting: System prompts, templates, few-shot
- L9 – Knowledge & Retrieval: Vector/graph indexes, rerankers, grounding, citations
- L10 – Tool & Function Invocation: Typed tool I/O (MCP), function calling, API bindings
- L11 – Addressing & Registry: Signed manifests, discovery, capability vectors, fingerprints
- L12 – Routing, Planning & Policy: Rule DSL + bandits, budgets, privacy, fallback/parallel
- L13 – Transport & Flow Semantics: Idempotent runs, streaming, CANCEL/RESUME, multiplex
- L14 – Session, Identity & Memory: Portable session envelope, capability tokens, memory tiers
- L15 – Governance, Safety & Schema: Redaction, validation/repair, schema change control, audit
- L16 – Application & Domain Logic: Product UX, workflows, agent frameworks
- Control: Policy/configuration management
- Management/Observability: Telemetry, evaluations, monitoring
- Security: mTLS, key management, PII protection
We don't claim to have all the answers. This proposal is intentionally incomplete and likely wrong in places. We're sharing it early because we believe the best ideas emerge from open discussion.
We're particularly interested in:
- Does this framing resonate or feel forced?
- What's missing or miscategorized?
- Are there better ways to think about these boundaries?
- What existing work should we learn from?
- 📄 AILIS Primer - Overview of the layer model
- 📋 Quick Reference - One-page summary
- 📝 All Proposals - Current and upcoming proposals
- 🔬 Reference Implementations - Code examples and tools
- 📊 Case Studies - Real-world system mappings
- 💬 Discussions - Community feedback and alternatives
This is an open invitation to think together about how we might better organize our understanding of AI systems. Whether you're building infrastructure, developing applications, or researching new approaches, your perspective would be valuable.
We follow an RFC-style process for proposals:
- 💡 Have an idea? Open an issue to discuss
- 📝 Ready to propose? Create a
draft/your-proposalbranch - 👀 Get feedback During 4-week review period
- ✅ Reach consensus Proposal accepted or declined with rationale
Quick ways to help:
- Share feedback - Use our issue templates
- Submit use cases - Show us real-world examples
- Propose alternatives - Challenge our assumptions
- Review proposals - Help evaluate new ideas
See CONTRIBUTING.md for the full process, and FEEDBACK.md for specific areas where we're seeking input.
This proposal emerges from practical experience building AI tools and observing the challenges of interoperability in the current ecosystem. We found ourselves wishing for a clearer map to understand how different capabilities relate to each other.
Rather than create Yet Another Stack Diagram™, we wondered: could we contribute something more broadly useful to the community?
- Documentation and specifications: CC-BY 4.0
- Code and examples: Apache 2.0
We chose these licenses to enable the widest possible collaboration and adoption, should any of these ideas prove useful.
Early Draft - This is version 0.1 of a proposal. Everything here is subject to change based on community feedback. We're not trying to create a standard—we're trying to start a useful conversation.
This proposal draws inspiration from:
- The OSI network model
- The work of countless AI infrastructure teams
- Open specifications like OpenAPI and GraphQL
- The Model Context Protocol (MCP) community
If you reference AILIS, please cite it via its DOI (see also CITATION.cff):
Darling, M. (2026). AILIS: A Proposed Layer Model for AI Systems (v0.2.0). Dollhouse Research. Zenodo. https://doi.org/10.5281/zenodo.21398506
Questions? Thoughts? We'd love to hear from you. Open an issue to start a discussion.