Give meaning
an identity.
Terms, definitions and relations become addressable structures. The goal is to traverse connected knowledge with its originating context preserved.
IL-LLMThe local-first
The local-first
LLM standard.
IL-LLM is Keddeh’s proposed standard for bringing language, knowledge and execution context closer to the local node.
Explore Keddeh SystemsGive meaning
Keep the working
context close.
Local node state, VFS resources and a seed ledger form the proposed working environment. Cloud independence depends on the particular engine and its actual dependencies.
Connect the layers.
Keep their roles.
IL-LLM handles semantic knowledge. KEX supplies state and addressing mechanics. BRAINK governs allowed execution. A module of KEX seeds is not interchangeable with any one layer.
Follow the construction path.
Word → node → node equipped with VFS → IL-LLM → bootstrapper → compiler → assembler → KEX seed → KEX engines → IL-LLM module of KEX seeds.
For each transition, identify the input, transformation, output, persistence and consuming component. Forward traversal should connect to its executable consequences; reverse traversal should recover the originating meaning and source.
What “standard” means here
This is Keddeh’s architecture and proposed local-first specification, not a claim of adoption by an external standards body. Reduced copying, local execution and performance must be reported per demonstrated implementation and workload.
UNDERSTAND THE ARCHITECTURE
IL-LLM. Knowledge with a place, a context and a path to action.
KEDDEH’s local-first intelligence concept begins with a different question: how should a system organise what it knows so that meaning, execution and evidence remain connected?
01 · The problem it addresses
A conventional application can distribute the same idea across documents, prompts, database rows and API calls. Each copy can lose its source or acquire a different meaning. IL-LLM proposes a local reality engine: a node maintains the knowledge, relationships and working state that matter to its own domain. Local-first describes where that context is held and processed; it does not mean every integration works without a network.
02 · Meaning is a structured relationship
A term is interpreted in context. In ClaimPath, “value” can mean original purchase price, current replacement estimate, insurer-approved allowance or a supplier offer. Those are separate attributes with separate evidence. An IL-LLM knowledge structure should retain the term, definition, domain, source, version and relationships rather than flatten them into one plausible answer.
03 · From words to addressable nodes
The proposed construction sequence is word → node → node equipped with a virtual file system → IL-LLM. A node gives a concept an identity. Its VFS holds the associated records or executable material. Relations describe how that node connects to other meanings and capabilities. Addressing identifies the record; it does not by itself prove that its contents are correct.
04 · Traversal produces an inspectable reasoning path
A request can be resolved from the local definition to relevant objects, evidence and available functions. For example, “find a replacement kettle” should traverse the recorded brand and model, the required capacity and electrical characteristics, source candidates and the customer’s selection. The output should expose the path and uncertainty, not silently convert a candidate into an approved purchase.
05 · Local reality and global agreement have different jobs
The whitepaper distinguishes the il-llm Local Reality Engine from the Ledger-LLM global consensus layer. Local work may evolve quickly. A shared record requires an explicit publication and acceptance process. This separation is useful because a local observation, a proposed update and globally accepted state are different stages. Network consensus also needs a specified membership, conflict and recovery protocol.
06 · BRAINK coordinates action; IL-LLM supplies context
IL-LLM provides the semantic and local-state context. BRAINK concerns actors, capabilities and coordinated work. A capability still needs a concrete handler, input contract and runtime. MCP makes such tools discoverable and callable; an SDK supplies application interfaces. Naming a tool or recording it in a workbook does not execute it. A successful operation should return an identifier and support a readback of what changed.
07 · KEX gives the transition an address and lineage
In the proposed construction chain, the bootstrapper, compiler and assembler connect the IL-LLM representation to KEX seeds and engines. KEX addressing and state geometry remain distinct from semantic meaning. A coordinate such as R121212_C121212 is a logical address until an implementation defines its mapping, bounds and storage behaviour. The one-origin convention is a design invariant to implement and test, not evidence that all collisions are impossible.
08 · The HTML carrier makes a working environment portable
The Omni-HTML concept packages interface, routing and resident application state in an HTML carrier. Depending on the application, JavaScript, a VFS and a separately supplied Python/WASM runtime can participate. ClaimPath’s personal carrier illustrates a bounded implementation: customer data and evidence can travel with an editable report. Network services still depend on reachable endpoints and the permissions of the user or runtime.
09 · Compaction is a storage contract
Law E describes a compact DNA seed in place of a massive logical allocation. To evaluate that mechanism, distinguish a digest, a compressed payload and a reference to external state. A hash proves consistency against matching data; it cannot reconstruct arbitrary missing data. A working seed format must identify what is retained, what can be regenerated and which dependencies must remain available.
10 · Constants need units, sources and acceptance rules
The whitepaper proposes a 0.297 phase-lock invariant and reference knowledge grounded in empirical constants. These are architectural proposals. A numerical threshold must define its units, measurement method, tolerance and failure behaviour. Scientific reference data needs source and version provenance. Shared constants alone do not establish cryptographic security, truth of an unrelated claim or agreement among competing nodes.
11 · Efficiencies should be demonstrated on the same task
Local context can aim to reduce repeated transfers, reconstruction of prompts and duplicate records. Explicit relations can aim to make lookup and audit more direct. Portable carriers can reduce application packaging overhead. Measure those hypotheses against an identified baseline: elapsed time, transferred bytes, peak memory, storage, completed task quality and recovery after failure. Publish the workload and configuration alongside the result. No universal speed or cost multiplier is asserted here.
12 · A practical learning exercise
Start with one item and give each attribute a source. Link the item to its photos, receipt and model label. Add two replacement candidates without overwriting the original. Record why one candidate is suitable, ask the user to choose, and preserve the resulting transition. This small exercise makes the architecture tangible: local context, typed relationships, bounded action and evidence of change.
Is IL-LLM a replacement model I can download and benchmark today?
This page explains KEDDEH’s architecture and its intended contracts. It is not a claim that a general-purpose pretrained replacement model, complete consensus network or universal deterministic compiler has been released. Use the software pages to inspect available applications and the research pages to examine evidence. Register your interest to follow implementation and learning material.
Learn the architecture, then evaluate a concrete mechanism. Keep the conceptual model, available code and measured runtime outcome distinct.