The compilation pipeline uses two models: an untrained model for pseudo-program generation and a trained compiler for LoRA extraction. Users should not need to know about this split. We also need version pinning for reproducibility.

Two-level naming following OpenAI's convention:

• Pretty alias (always points to latest): paw-4b-qwen3-0.6b
• Exact snapshot (immutable, dated): paw-4b-qwen3-0.6b-20260325

The backend maintains a registry mapping each name to the full pipeline config (pseudo-gen model, trained compiler path, LoRA mapper path, compatible interpreters, etc.).

paw-{compiler_size}-{interpreter_model}-{date}
     4b               qwen3-0.6b        20260325

• Users see a simple name; the two-model split is hidden
• Alias updates silently when we train better models
• Power users can pin to a snapshot for reproducibility
• .paw files always store the exact snapshot name
• Backend validates compiler-interpreter compatibility at request time
• Adding new compiler/interpreter combos is just a registry entry