The LLM produces a plan (JSON tree) and we render it to UVL deterministically, so the feature tree is guaranteed syntactically valid.
Constraints (if enabled below) go through a second LLM call and are validated one-by-one — any that break the model are dropped.
2. Models
Model
Size
Status
Actions
3. Generate
{# Generation-scoped progress bar. The one in the Models card
footer still exists for model downloads; this mirror
appears right above the results so users see progress
while variants stream in — same pattern as the random
generator's step 6. #}
{% endblock %}
{% block scripts %}
{# The generator's bundles, in this order: icons and tooltips (ui.js); the
Pyodide runtime, the same as the random generator's, which validates
every candidate with flamapy; CodeMirror and the UVL mode of the "View"
dialog; and the page's controller, which drives the seed picker and the
WebLLM worker and reuses window.generatorRuntime. #}
{{ feature_script('generator', 'ui') }}
{{ feature_script('generator') }}
{% include "generator/_codemirror_uvl.html" %}
{{ feature_script('generator', 'llm_page') }}
{% endblock %}