model-atoms

AI model definitions canonicalized — Claude, GPT, Llama, Gemini, Mistral variants with capabilities, context windows, pricing tiers, deprecation policy, tool-use shape, vision support, modality. One typed source-of-truth for what a model is, what it can do, and when it goes away.

model-atoms is a *-Atoms catalog in the Convergent Systems ecosystem. It defines what exists in its domain — typed, versioned, machine-readable, composable, and open — so runtimes (and humans) can stand on shared infrastructure instead of reinventing it.

Structure

model-atoms/
├── ATOMS.yml              # Catalog manifest
├── atoms/                 # Reusable building blocks
├── models/                # Compositions assembled from atoms
├── rules/                 # Typed constraint vocabulary
├── schemas/               # Catalog-specific JSON Schemas
├── exports/               # CI-generated machine-readable exports
└── docs/                  # Human-readable documentation

Atom types

• model-card — A specific model variant (claude-opus-4-7, gpt-4o, llama-3.1-70b-instruct) with vendor, family, and reference identifiers.
• capability — Discrete capability (tool-use, vision, function-calling, json-mode, streaming, prompt-caching, extended-thinking).
• pricing-tier — Input / output / cache-hit / cache-write rates with currency and effective date.
• deprecation-policy — End-of-life shape (announced date, replacement model, hard stop date).
• tool-use-shape — Vendor's tool-use API shape (parallel function calls, XML tool tags, JSON tool blocks).
• modality — Input/output modality (text, image, audio, video, structured-output).

Rule types

• model-compatibility — Which atoms a model supports (e.g., gpt-4o supports parallel tool calls; llama 3.1 does not).
• capability-grant — How a capability is enabled (request flag, API version, account tier).
• deprecation-window — Time between announcement and hard stop (e.g., Anthropic: 6 months minimum).

Runtime consumers

aish, olympus

How to consume

Machine-readable exports are published in exports/ on every release:

• exports/catalog.json — full catalog dump (every atom, composition, rule)

Exports are deterministic, signed, and versioned. See ATOMS.yml for the manifest and the conformance spec.

How to contribute

1. Read ATOMS.yml to understand the catalog's atom types, compositions, and rules.
2. Add a new atom under atoms/<type>/ or a composition under models/<name>/.
3. Open a PR. CI validates the schema, references, and exports.
4. Larger structural changes go through the XAIP process.

Ecosystem

• Federation: xdao.co · github.com/convergent-systems-co/xdao
• Spec: github.com/convergent-systems-co/atoms-spec
• Tools: github.com/convergent-systems-co/atoms-tools
• Umbrella: github.com/convergent-systems-co/atoms — all catalogs as submodules

License

Apache-2.0 — see LICENSE.