Legacy observations

Methodology-v0 OBSERVED peak VRAM may appear on a model page only when a legacy benchmark row maps cleanly to that model.

Known methodology-v0 limitations include:

• configured context is not actual prompt length
• runs were often single-shot
• process VRAM collection may be unreliable
• total VRAM may include other GPU users
• no baseline VRAM subtraction
• exact llama.cpp commit/build may be missing
• artifact hash may be missing

No mapped legacy observations are currently imported for the curated models. Pages omit OBSERVED peak VRAM rather than inventing values.