Warning
Restart planned following a development hiatus. I intend to resume work on the full ELAI project. This repository currently contains unfinished research code. Historical implementation, security and benchmark claims have not been revalidated. Do not expect a working product. This status update announces that commitment; it does not establish new functionality, completed fixes or a release date.
ELAI explored a local-first, model-agnostic agent harness: models propose actions, while software owns planning, state, permissions, execution, evidence, verification and the decision that work is complete. That was the design goal; it is not a claim that every path enforced it correctly.
I used it to try too many ideas headlessly, not as software I relied on for actual work. Retrieval, orchestration, verification, sandboxing, memory, model routing, self-improvement and UI experiments kept accumulating. It became overengineered, and I stopped development at that time. I now intend to return to the project. The lesson is simple: build software you want to use that solves your problems, not all the problems you know about.
- Evidence outside the generating agent: judge work through independently captured artifacts.
- Fail-closed verification: reject missing evidence instead of treating it as a skipped check.
- Journals and replay: preserve what happened and explore reproducible execution.
- Bounded workers: keep task state, permissions and completion decisions in the harness.
- Code retrieval: combine keyword/vector search, dependency graphs and repository maps within a context budget.
These are extraction candidates and design references. The salvage map links to source code, integration points and concrete limitations.
| Material | What you can conclude |
|---|---|
| Rust, Python and Tauri source | Implementation exists; current buildability and end-to-end behavior are unverified. |
| Benchmarks and recorded results | Harnesses, fixtures and historical result artifacts are included; their scope and missing inputs matter. |
| 659-entry feature catalog | These are plan units and experiments, not 659 working features. |
The archive preserves 883 real historical commits, with privacy edits to their contents and identities, followed by archive-preparation commits. Dates and parent relationships were retained; see the history review.
- Postmortem — why the project expanded and what I would do differently.
- Salvage map — concrete places to start reading or extracting.
- Benchmark guide — what was committed and what the results actually cover.
- Conductor — the external automation that coordinated ELAI's development, with a guide to the selected source.
- Detailed research overview — the former long README, including the full catalog.
- Recovered documentation — dashboard designs, mockups, research sweeps, experiment reports and workflow guides.
- Systems specification — the intended architecture.
- Archive guide and publication review — selection, privacy edits and limitations.
Older files also call the project LIA. Read, fork or extract whatever is useful, expecting to do your own validation. Original code and documents use MIT; third-party notices cover retained upstream material. I plan to resume development and maintenance across the project. Release timing and support availability are not yet established.
