Sunstone North Lab · all-HTML version
Same head, different silicon
The 22 MB Sunstone retrieval head was trained exactly once: bf16 precision, datacenter CUDA, a 31B host. Then everything a deployment varies was varied. Numbers are the published release-protocol results from the SRT repository and paper §11.6.4.
Why one matrix is enough
| Bridge | Image → text R@1 | Note |
|---|---|---|
| Centered cosine (zero training) | 0.288 | the free baseline |
| Orthogonal Procrustes | 0.226 | refuted: worse than nothing |
| One trained linear matrix | 0.661 | 22 MB, 117k pairs |
| Two-layer MLP | 0.567 | 33× the data, never wins |
On the literature-standard Karpathy 5k split the same head scores 0.416 R@1, matching fully trained 2018 dual encoders from roughly 3,000 times less pair data. Every rung carries a shuffled-pairs control that collapses to chance.
Change the host: 31B → 3B
The full ladder was re-run on Qwen2.5-VL-3B, ten times smaller. Same anisotropic-linear fingerprint. At an identical 39,000-pair budget the small host's head lands at 0.577 R@1, inside the big host's 0.553–0.590 range. Ten times less model, nothing lost.
Change the precision: bf16 → 4-bit
The bf16-trained head, applied unchanged to 4-bit NF4 states: −0.01 R@1. A 42 KB mean recalibration, a single vector, recovers half of that.
Change the silicon: datacenter → the chip in a Mac
The same images embedded on datacenter CUDA and on Apple Silicon, different kernels, different quantization. Through the head, every local state retrieved its datacenter twin at rank one: 100% R@1 agreement.
10× host reduction with no loss · −0.01 R@1 at 4-bit with unchanged weights · 100% cross-silicon agreement · one artifact, every tier. Train once, read everywhere.