Moonshot
Train the model under randomized artifact bottlenecks, not just weight noise.
Examples:
- some steps lose protected subsets
- some force harsher codebook collapse
- some remove exception paths entirely
- some enforce extra sharing or shallower reconstruction
The goal is to make the model robust to the kinds of mutilation that real artifact compression causes.
Why this is outside the current prior
Standard noise injection and quantization-aware training usually simulate one degradation family. Artifact dropout simulates whole failure regimes of the final submission object.
Mechanism sketch
During training or late-stage adaptation, randomly sample artifact constraints such as:
- no protected rows this step
- reduced codebook capacity this step
- stricter shared-depth regime this step
- harsher clipping or coarser packing this step
The model learns to distribute competence so it does not rely too heavily on one fragile artifact path.
Why it might matter for Parameter Golf
Parameter Golf cares about surviving the exact compressed route, not looking elegant before export. Artifact dropout directly trains against fragility to that route.
Cheapest falsifier
- simulate two or three compression failure modes during a short finishing phase
- compare post-roundtrip robustness versus a standard finishing run
Kill it if robustness broadens but best-case final artifact quality falls too much.
What would make it real
- narrower pre→post degradation gap
- better robustness to export/config changes
- survival under multiple compression paths without massive quality sacrifice