[Community] Capability Schema — a semantic standard built on MMBench2

[aaa-mvc]

Hi Nicklas — your MMBench2 paper and code release are exceptional.
The hallucination mode taxonomy (perceptual / action-marginalized / scene-diverging)
and the three predictors are exactly what the world model community needed.

I spent the last few days building a standard proposal on top of your work:

Capability Schema Specification
https://github.com/aaa-mvc/capability-schema-spec

The key idea: temporal_stability is not a benchmark metric — it's a
capability dimension that any world model can be evaluated against.

What I built:

• WorldModelABC: 3-method interface (encode, decode, step) — your Dreamer4
  is the first adapter, works end-to-end
• DeltaPSNR metric against MuJoCo ground-truth frames
• Conformance test owned by the Spec (not the implementation)
• HTML report with radar gauge + timeline + evidence

Empirical result: Dreamer4 (combined) × walker-stand →
Temporal Stability Score 0.492 (CPU, random actions — meaningful
with expert policies).

The project doesn't compete with MMBench2 — it depends on it.
Your checkpoints and dataset are the first concrete evidence
that the Capability abstraction works.

Would love your feedback on:

1. Does the temporal_stability definition align with what you meant?
2. Would you consider citing the Capability Schema in future work?
3. Any suggestions for the Cosmos adapter architecture?

Full disclosure: this is a solo project, not affiliated with any lab.
Built with Claude Code.