model-specs

MODEL SPECS — this skill is knowledge-only: no cluster calls, no scaling/inference actions. It's the reference for what a catalog model *is*, as opposed to the models skill (what a deployed instance is currently doing). ## Scripts Run via the shell tool: `python3 "$SKILLS_ROOT/model-specs/scripts/get_model_reference.py" --model-name <name>`. Each catalog model with a documented dataset/training profile has a file under `references/<model_name>.md` (e.g. `references/pi05.md`, `references/dreamzero.md`), keyed by the same catalog directory name used elsewhere (platform/base/models/<model_name>/). A model with no file here either has no fine-tuning recipe on this platform or hasn't been documented yet — don't guess its specifics from general knowledge of the base model. pi0.5's reference is a Dimension/Priority checklist table matching the datasets skill's own DATASET COMPATIBILITY CHECKLIST row for row — read them side by side by row number when validating a candidate dataset. DreamZero's is plain training-data provenance prose instead of a table, since it's inference-only on this platform (no recipe to check a dataset against).