Explore the layer evidence
Loading a recorded model run…
Voyager decision
—
- Routed score
- —
- Fixed ID threshold
- —
- Margin
- —
Final-layer GL-MCM
Scores and margins are shown ×1,000 for readability. Higher scores are more ID-like. These are not OOD probabilities.
Which layers support the score?
Each contribution is the layer’s ensemble weight × (its score − the fixed threshold).
Local ID matches
Actual 224 × 224 CLIP crop. Colour is local maximum ID-class match within this layer, not attention or causal importance. Map scales vary across layers.
Five hard expert routes
Each router selects one subset. The final layer is always included; selected scores are averaged across routers.
| Router | Selected layers | Score ×1,000 |
|---|
Closest ID concepts
- Global MCM ×1,000
- Local MCM ×1,000
- Global + 0.5 × local
- Layer’s ensemble weight
- ID-reference percentile
Cosine similarities describe matches to the ImageNet label set. A plausible match can also occur for an OOD image.
How to read this explanation
Decision scope. “ID-like” means above a fixed threshold relative to ImageNet-1K, not that the image is known, safe, or correctly classified. “OOD-like” means below it. The threshold is the fifth percentile of 50,000 cached ImageNet validation ID scores; 95% retention applies to that reference set, not every new image.
Selected examples. ID/OOD reference labels describe membership in the ImageNet-ID versus OOD benchmark datasets. OOD datasets can contain familiar ImageNet objects. Cases were chosen at fixed thresholds to show contrasting outcomes, including a failure; they do not estimate aggregate performance. Every published crop was rerun through the trained model.
Exact arithmetic. A layer’s weight is its membership in the five selected experts, divided by each expert’s subset size and by five. Weights sum to one. Weighted layer scores sum to the final score; centering each term on the same threshold sums to the final margin. The signs describe this decomposition, not independently calibrated layer decisions.
Interpretation limits. This trace shows the computation the method actually used. The local map is built from attention value-token ID matches, without attention weights. It is not a causal explanation and does not prove which pixels caused the model’s route. Unselected layer features still enter the router’s mean-pooled input.