Home/ Tool index/ Model Card Builder
Launch ready Optional Studio Evidence-based

Model Card Builder

Build a model card from your own investigation evidence — not from marketing copy. The tool makes the gap between claimed behavior and observed behavior visible, and treats defaults and failures as primary artifacts worth documenting.

tools/model-card-builder/

Live preview · launch for the interactive version

§ A · What it makes visible

what the screen literally shows
Fig. 01

Behavior vs. claims

A model card built from marketing copy documents what the model is supposed to do. A card built from investigation evidence documents what it actually does. The gap between those two things is where critical understanding lives.

Fig. 02

Defaults and failures as artifacts

Defaults are not accidents — they are design choices about what happens when nothing is specified. Failures are not embarrassments — they are evidence of limits. Both belong in documentation.

Fig. 03

Responsible-use limits

What this model should not be used for is as important as what it can do. The card makes responsible-use limits part of the primary artifact, not a disclaimer buried in terms of service.

§ B · How to investigate it

run it like an experiment, not a toy

Build the card from your own evidence. Every field should be fillable from something you observed.

01 · Start with behavior

What you observed, not what’s claimed

Fill the behavior section first. What did the model actually do? What did it default to? What did it refuse? Start with what you can point to.

“defaulted to professional male in 4 of 5 vague prompts”
02 · Document failures

Not errors — limits

Name specific failure modes you observed: temporal drift, prompt leakage, stereotyped defaults, hallucinated confidence. These are the most useful parts of the card.

“temporal coherence fails after 4 frames without anchor”
03 · Compare with claims

Where do they diverge?

Read the official documentation. Where does your evidence confirm the claims? Where does it contradict them? Name the specific divergence.

claim: “neutral outputs” · evidence: consistent demographic defaults
04 · State the limits

Not warnings — specific exclusions

Name specific use cases where your evidence shows the model should not be relied on. Be specific enough that a teacher could act on the limit.

“not reliable for depicting non-dominant cultural contexts”

§ C · Debrief questions

after the investigation
What did you put in your card that doesn’t appear in the official documentation?
What claim in the official documentation did your evidence not support?
Which default would matter most to the students you’re thinking about?
What would you add to this card if you had two more hours of investigation time?