From intent
to action.
By design.

Eyleme is a Rust platform in development for composing bounded agentic jobs from model roles, tools, reusable skills and project knowledge. Humans define goals and authority; the platform records actual outcomes. Its implemented local task foundation is distinct from the broader model, RAG, memory and learning design.

The core includes the task console and scoped native loop/model contracts, project-data gates, explicit-reference memory context, typed counter tools, reusable counter skills and decision evidence. Each component has its own proof and limits. This promotional site does not run them, and those proofs do not establish general inference, graph search or complete harness jobs.

Yes, in a bounded native macOS counter demonstration: an actual device-owner check, exact approval, effective stop before the remaining action, revocation and state-preserving reopen passed on 14 October 2026. This grants no data-training or sharing rights. The separate start --human/browser flow, remote identity and arbitrary model jobs were not part of that acceptance.

No. The optional roles are perception, reasoning, planning, decision, action policy, world model, specialist, critic and representation/retrieval. One model may fill several compatible roles. Audio, vision and text profiles are distinct, and exact model acceptance is required before execution.

Declared model artifact and role contracts are implemented; they do not load weights or run inference. Local assignments still report unsupported execution. The architecture permits explicitly configured external frontier models, but configuration is not authorization to disclose private data or make paid calls. A local-only profile never silently falls back to a hosted model.

The design grounds answers in permitted project documents and optional working, episodic, semantic, procedural, knowledge-graph and analytical memory. The current Rust adapter builds bounded context from explicit source references, rechecking rights, expiry and lineage. Semantic/vector retrieval, graph traversal, analytical engines and tokenizer counting remain separate acceptance gates.

Native decision evidence joins submitted alternatives, constraints and declared model pins to actual scoped tool receipts and complete, partial or unknown outcomes. Confidence and model declarations remain unverified inputs. The current proof uses deterministic counter tools; model explanations, causal claims and whole-loop performance require separate evidence.

Not automatically. Execute, collect, train, external disclosure and ecosystem export are independent purposes. Derived records inherit current source restrictions; withdrawal blocks future use but does not erase published copies or unlearn weights. Ecosystem packages need explicit selection, provenance, terms, review and compatible import acceptance.

The plan accepts training per exact model: owned Rust gradients, CPU/Metal LoRA and separately supported QLoRA, followed by independent evaluation and controlled promotion. Unsloth-like tooling remains a reviewed reference or recipe aid with separate export/reacceptance gates. No accepted native training or improved-model result is claimed.

The first experimental target is a single-user loopback application on a MacBook Pro with Apple M4 Pro and 24 GB unified memory. Browser and native execution have distinct boundaries. Mac mini M4 and Linux RTX equipment are additional inventory, not accepted support; remote access and Linux acceptance are future scope.

Eyleme first-party core code and documentation use Apache-2.0. Separately identified third-party material, model weights and datasets retain their own terms. This does not grant a blanket license to every asset or model, or change the website repository license.