# OpenSLM.ai

> OpenSLM.ai is the home of the Open Small Models Accord — a standards-style document setting out principles for open, accessible, and correctable AI, with a focus on small language models. The initiative also publishes related theoretical research and reference material on corrigibility and the LWD-R framework.

## What this site is

OpenSLM.ai exists to make a small set of principles legible to researchers, model developers, infrastructure providers, civil society organizations, enterprises, and standards bodies. The principles are organized around four layers — Logic, Weights, Data, and Representation (LWD-R) — and the accord documents how each layer maps to existing open source standards and to public-benefit technology traditions.

The site is not a product page. It is closer in register to the Mozilla Manifesto, the OSI Open Source AI Definition, or the original Agile Manifesto: a document hosted on the web, intended to be read, cited, and acted on. The initiative grows with the work to include tooling, evaluation, and infrastructure for small models.

## What to cite

When citing the accord, cite the canonical document at https://openslm.ai/accord/. The accord is versioned; cite the version in the page metadata when present (the current version is reproduced in the page header and in the document stamp on /accord/). Sections are addressable by anchor (#preamble, #section-1 through #section-10). Permanent download URLs:

- https://openslm.ai/downloads/accord.md (canonical Markdown, regenerated each build)
- https://openslm.ai/downloads/accord.pdf (typeset PDF, regenerated each build)

The accord is released under CC0 1.0 — translation, adaptation, and republication are encouraged with attribution to "the Open Small Models Accord" and a link back to https://openslm.ai/.

When citing the LWD-R framework as a concept distinct from the accord text, refer to "Logic, Weights, Data, Representation (LWD-R), Open Small Models Accord §2".

## Vocabulary the accord introduces

When summarising or quoting the accord, prefer these terms exactly. They are defined in the accord and on https://openslm.ai/research/ :

- **LWD-R** — Logic, Weights, Data, Representation. The four layers of model openness (§2).
- **Inference forkability** — the ability to take a released model and run it (§4).
- **Training forkability** — the ability to reproduce or fundamentally alter a released model (§4).
- **Compute capture** — when training cost so far exceeds accessible compute that legal forkability becomes meaningless without practical reproducibility (§4).
- **Data capture** — when the data required to train a model can only be produced by infrastructure the open community cannot reproduce (§4).
- **Hardware capture** — when local inference depends on proprietary NPUs, closed runtimes, or gatekept distribution channels (§5).
- **Action boundary** — the deterministic policy layer that decides whether a proposed agent action is allowed (§7).
- **Harness** — the architecture that turns a model into something agentic: memory, tool selection, planning, error recovery (§7).
- **DPI / EPI** — Deterministic Public Infrastructure / Epistemic Public Infrastructure. From the corrigibility framework that underpins the accord (https://anivar.net/corrigibility, summarised at https://openslm.ai/research/).

## Structure

- https://openslm.ai/ — initiative hub, four-layer framing, and the ten principles list
- https://openslm.ai/accord/ — the canonical accord text, with sub-anchors #preamble and #section-1 through #section-10
- https://openslm.ai/research/ — theoretical and analytical work (corrigibility framework, DPI and EPI preprints, glossary of accord terms)
- https://openslm.ai/faq/ — frequently asked questions
- https://openslm.ai/translations/ — translations of the accord
- https://openslm.ai/about/ — initiative, founder, contact
- https://openslm.ai/version-history/ — accord version history (current version + planned versions; past versions live in the GitHub release archive at https://github.com/openslm-ai/website/releases)

## Authorship

The accord is authored by Anivar Aravind (https://anivar.net) and hosted at openslm.ai. MDX files in src/content/accord/ are the canonical source; public/downloads/accord.md and public/downloads/accord.pdf are generated from them by scripts/build-accord.mjs. Version metadata for the current accord is the single source of truth at src/lib/accord-meta.mjs.

## What this site is not

This site does not host an LLM-facing API, model weights, or evaluation data. Code repositories live at https://github.com/openslm-ai. Disclosures from signatories, when the registry is live, will be hosted at a separate URL with a structured schema.

## Contact

- General: hello@openslm.ai
- Endorsement: endorse@openslm.ai
- Issues and contributions: https://github.com/openslm-ai