MODEL / BEFORE THE FIRST TOKEN

From weights
to a world you can inspect.

Admit the source. Build its container. Follow the structure.

01SOURCE→02PLAN→03ENCODE→04INSPECT→05OBSERVE

VINDEX3 / EXPLICIT GENERATION

01 / SOURCE CHECKPOINT

Start with the model.

Sources · one artifact per line

Hugging Face references or paths on the runner. Planning an hf:// source reads headers and metadata. Encoding reads the bound tensor ranges.

Admission scope

Text admission requires understood semantics, present operands, and execution support. It is distinct from whole-model completeness.

New output container · runner filesystem

Connect a runner that advertises planning for these source forms, or import a CLI plan.

Use the CLI without a connected runner

larql vindex3 plan 'hf://google/gemma-3-4b-it' --output plan.json

Open plan.json here afterwards. This does not execute or upload the command.