Moonshot
Stop asking only which weights are outliers. Ask whether outliers and protected regions live on a compressible topology:
- shared columns
- repeated channels
- bands
- motifs
- blockwise families
- masks reused across layers
Why this is outside the current prior
Most selective-precision work hunts the best ranking. This moonshot says a slightly worse ranking with a far more compressible shape may beat the mathematically best but irregular exception set.
Mechanism sketch
- search for structured protected subsets
- bias exception maps toward reusable topologies
- measure byte cost after encoding the exception layout itself
- allow only a small number of exception families shared across tensors
Why it might matter for Parameter Golf
At hard artifact caps, metadata becomes first-class. If the protected subset is irregular, the side information can erase the gain. If the subset has reusable structure, the same budget can buy much more protected fidelity.
Cheapest falsifier
- compare unstructured top-k protection against structured masks at equal final encoded bytes
- measure whether structured masks lose too much raw quality before the codec savings make up for it
Kill it if structure costs too much quality and never wins back enough bytes.
What would make it real
- structured masks beat unstructured protection at equal final artifact size
- exception families transfer across layers or repeated blocks
- the encoded layout cost becomes dramatically smaller than arbitrary sparse masks