AILIS Cheat Sheet¶
A scannable field guide for the AI Layer Interface Specification.
AILIS is a proposed map for discussing AI systems. Use this page when you need to orient quickly, map a product, or ask which layer a concern belongs to. For details and examples, open the Layer Atlas.
The stack in five regions¶
- L0-L2Infrastructure FoundationFacilities, compute fabric, driver/runtime substrate.
- L3-L7Model and Inference StackGraphs, numbers, encoders, weights, serving.
- L8-L10AI Application InterfacePrompts, retrieval, tools, typed invocation.
- L11-L15Orchestration LayersRegistry, routing, flow, session, governance.
- L16+Application and Domain LogicProduct workflow, domain behavior, user experience.
Layer field card¶
- L16Application & Domain LogicProduct UX, workflows, domain rules.
- L15Governance, Safety & SchemaValidation, redaction, guardrails, audit.
- L14Session, Identity & MemoryPortable sessions, consent, capability tokens.
- L13Transport & Flow SemanticsStreaming, cancel/resume, idempotency.
- L12Routing, Planning & PolicyProvider choice, ensembles, budgets, privacy rules.
- L11Addressing & RegistryDiscovery, manifests, fingerprints, capability metadata.
- L10Tool & Function InvocationMCP, function calling, APIs, tool schemas.
- L9Knowledge & RetrievalVector and graph indexes, reranking, citations.
- L8Context Construction & PromptingSystem prompts, templates, context packing.
- L7Inference Engine & DecodingServing, batching, KV cache, token streaming.
- L6Model Parameters & ArchitectureWeights, adapters, model families.
- L5Tokenization & EncodersText tokens, image/audio encoders, embeddings.
- L4Numeric & QuantizationPrecision, sparsity, compression, calibration.
- L3ML Graph & CompilationIR, lowering, optimization, execution targets.
- L2System & Driver RuntimeCUDA, ROCm, Metal, device memory.
- L1Compute FabricGPUs, TPUs, NPUs, CPUs, memory, interconnect.
- L0Facilities & PowerDatacenters, power, cooling, physical security.
Fast mapping questions¶
When evaluating a project, ask:
| Question | Why it matters |
|---|---|
| What is the primary layer? | Prevents "platform" or "agent" from hiding the real function. |
| Which adjacent layers does it include? | Finds bundled assumptions and integration risk. |
| What handoff does it define? | Clarifies data, control, identity, policy, or trust boundaries. |
| What would break if this layer were swapped out? | Reveals lock-in and missing interfaces. |
| Is this solving a layer problem or a product problem? | Keeps infrastructure research distinct from application UX. |
Under-served middle¶
AILIS currently treats L11-L15 as the most important open area:
| Layer | Core question |
|---|---|
| L11 Addressing & Registry | What exists, which version is it, and what evidence supports its claims? |
| L12 Routing, Planning & Policy | Which model, tool, provider, or ensemble should run under the current constraints? |
| L13 Transport & Flow Semantics | Can work stream, pause, cancel, resume, retry, and report progress reliably? |
| L14 Session, Identity & Memory | Who is acting, what context is portable, and what memory may be used? |
| L15 Governance, Safety & Schema | What must be redacted, validated, repaired, approved, logged, or blocked? |
Core artifacts¶
| Artifact | Likely layer |
|---|---|
| Signed model, tool, prompt, or index manifest | L11 |
| Capability vector or fingerprint | L11-L12 |
| Routing policy or planner graph | L12 |
| Resume token or idempotency key | L13 |
| Portable session envelope | L14 |
| Memory reference with consent metadata | L14-L15 |
| Output schema, validator, or repair loop | L15 |
| Audit log tying prompt, model, tool, identity, and output together | L14-L15 |