Capability Schema Specification

Capability Schema does not replace world model benchmarks. It gives them a shared language.

---

What is Capability Schema?

Capability Schema defines a standard way to describe, observe, and report the capabilities of generative world models.

This is not a benchmark

Benchmarks produce a number. Capability Schema produces a semantic unit — a Capability Score that is comparable across models, queryable, and traceable to evidence.

One Capability, One File

capabilities/
  └── temporal_stability.yaml

Each file defines:

• What the Capability means
• What metrics observe it
• What predictors estimate it (when ground-truth is unavailable)
• How to report it

Model-Independent

Dreamer4. Cosmos. Genie. Any future world model.

Same Capability definition. Different adapters.

---

Quick Start

# Install the reference implementation
pip install capability-schema

# Or use the spec directly — Capability Schema is a document standard,
# not a software dependency. Anyone can implement it.

See capability-schema-reference for the reference implementation.

---

Repository Structure

capability-schema-spec/
├── capabilities/           ← Capability definitions (YAML)
├── conformance/            ← Conformance test suite
├── rfcs/                   ← Historical RFCs and proposals
├── CHARTER.md              ← Governance charter
├── GOVERNANCE.md           ← Governance rules
├── PROPOSAL.md             ← Project establishment report
└── README.md               ← This file

---

Governance

See GOVERNANCE.md. Three constitutional axioms. BDFL → Editorial Board → Foundation.

---

Conformance

• Anyone may claim they use Capability Schema.
• Only those who pass the Conformance Test may claim "Capability Schema Conformant."

---

How to Cite

We report results using the Capability Schema Specification.

Short form: "following the Capability Schema"

---

Built On

Capability Schema is built on the work of others:

• MMBench2 (Hallucination in World Models is Predictable and Preventable) by Nicklas Hansen & Xiaolong Wang (UC San Diego) — the dataset, checkpoints, and hallucination taxonomy that first identified Temporal Stability as a measurable dimension. The Reference Implementation wraps MMBench2's pretrained Dreamer4 model as its first WorldModel adapter.

---

Runtime Implementations

The Capability Schema is a document standard. These projects consume it as a runtime layer:

Project	Role	Link
TraceOS	AI experiment operating system — uses Capability Schema YAML as input to its Analytics Engine via a transformation-only adapter	aaa-mvc/traceos
capability-schema-reference	Reference implementation wrapping Dreamer4 + MMBench2	aaa-mvc/capability-schema-reference

To add your project, open a PR adding a row to this table.

---

License

• Specification: CC BY 4.0
• Conformance Tests: MIT