Docs
Reading a model card
A Model Card is the smallest readable rating unit: one card = one world-model version × use-domain, always paired with evidence strength and run conditions.
Model card anatomy
Callouts for use-domain, level, evidence strength, seven dimensions, and run conditions.
MODEL CARD
Prediction: driving & robotics
Example-WM · 2026.08
Level
2
Evidence strength
Seven dims (sketch)
Run conditions
N=1 · bf16 · RTX 4090
- 01
Use-domain
The conclusion holds only inside this domain.
- 02
Level
1 / 2 / 3 / NR: maturity, not a leaderboard place.
- 03
Evidence strength
High / medium / low: how strong the evidence is—paired with level.
- 04
Seven scores
Shown separately; never collapsed to one total.
- 05
Run conditions
N, precision, GPU, …—the trust boundary on the card.
Start with use-domain and model version—without them, numbers mean nothing. Then check that level and evidence strength appear together.
The seven-dimension radar or table is detail, not a total. Deployers should watch uncertainty calibration and failure boundary closely.
Run conditions answer under which hardware, precision, and sample regime the number was obtained. With a subset or N=1, do not over-read strength.