lora-kernel
llms.txt4 starsChanged 58 days agoApache-2.0View on GitHub
Experts defined by their corpora: one small LoRA per job on Gemma 4 E4B, trained on how to navigate an editable markdown knowledge base (search, open, follow links, calculate) rather than on its facts; a very small router that abstains to a frontier model outside every expert's corpus; served as an OpenAI-compatible API with OpenClaw on top. Apache-2.0. Every claim is marked [ran] (observed, run directory named), [read] (from source or a paper) or [spec] (decided, not built). Measurements are…
In this repository2 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- experiment-briefSkill
The file
Files
# lora-kernel > Experts defined by their corpora: one small LoRA per job on Gemma 4 E4B, trained on how to navigate an editable markdown knowledge base (search, open, follow links, calculate) rather than on its facts; a very small router that abstains to a frontier model outside every expert's corpus; served as an OpenAI-compatible API with OpenClaw on top. Apache-2.0. Every claim is marked [ran] (observed, run directory named), [read] (from source or a paper) or [spec] (decided, not built). Measurements are on generated suites, synthetic organisations and real regulations ingested verbatim; no real user traffic yet. Key results, all [ran]: an inbox-triage expert 0.989 vs the bare base's 0.345; multi-hop questions over an unseen wiki of atomic statements, 38/40 on Gemma 4 E4B (bare 19/40); the same fluids adapter 11/90 through `tool_calls` and 90/90 served with results inline as its corpus taught — serve an expert the way it was trained before concluding a capability is absent. Not solved: multi-hop citation on real documents (13/21; four treatments tied their baseline); the router that passes (ROUTE0, a factored rule: 0 of 600 foreign texts served locally) is literal — paraphrases leave for the frontier, and no learned router passed; small models still invent, caught by a grounding check outside the model; no member yet scored on an external benchmark (τ²-bench installed, T0). ## Start here - [README](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/README.md): the shape, the five pieces of 1.0, and a dated status: what works, what does not, what is not implemented - [Memory specification](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/docs/MEMORY.md): library, radar, three verbs (`<search>`, `<open>`, `<calc>`), the navigating LoRA, the runtime referee - [Record](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/docs/RECORD.md): everything measured, including negative results and instruments that lied, each line naming its run - [Plan](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/docs/PLAN.md): living milestones, each with gate and falsification condition written before it ran ## Weights - [distributor-wiki@v2 on Hugging Face](https://huggingface.co/Matias/lora-kernel-distributor-wiki-gemma4-e4b): the trajectory LoRA on gemma-4-E4B-it, research preview, with its limits - [Colab notebook](https://colab.research.google.com/github/EvolvingAgentsLabs/lora-kernel/blob/main/examples/colab/wiki_walk.ipynb): bare vs adapter on multi-hop questions, then edit one note and ask again ## Architecture and serving - [Architecture](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/docs/ARCHITECTURE.md): experts, router, memory, runtime, the small/large pair, the frontier - [Serving](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/docs/SERVING.md): vLLM multi-LoRA behind the OpenAI-compatible proxy (`training/harness/openai_proxy.py`) - [OpenClaw](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/docs/OPENCLAW.md): agents per task on top, as run live - [Demo](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/docs/DEMO.md): a reference organisation (school, distributor) with signed identity, held payments, grounding - [Examples](https://github.com/EvolvingAgentsLabs/lora-kernel/tree/main/examples): code-only reference organisations and MCP servers ## Why and how it is measured - [Knowledge trajectories](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/docs/KNOWLEDGE-TRAJECTORIES.md): why weights should hold navigation and notes hold content; self-contained, written for review by other models - [Framework gaps](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/docs/FRAMEWORK.md): what works, what does not, what a generic framework would still need - [Foundations](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/docs/FOUNDATIONS.md): the mathematics, each section tied to the run that instantiates it - [Instructions for coding agents](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/CLAUDE.md): the measurement rules this project paid for ## Optional - [Article: an organisation that runs on agents, on a single GPU](https://github.com/EvolvingAgentsLabs/lora-kernel/blob/main/docs/articles/2026-09-it-was-the-harness.md) - [Spanish documentation](https://github.com/EvolvingAgentsLabs/lora-kernel/tree/main/docs/es) - [Record before 2026-09-19, tag v0.1-foundations](https://github.com/EvolvingAgentsLabs/lora-kernel/tree/v0.1-foundations)
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

