Inspect the computation behind a generation
llm-scry explores an interface for moving from a generated token to the model state behind it. The implemented backend loads a TransformerLens model, streams tokens and alternative-token log probabilities, then captures attention patterns and residual, attention-output, and MLP-output activations in a final forward pass.
Session endpoints expose generated tokens and attention patterns by layer and head. TransformerLens supplies the model hooks; this project supplies the API and session plumbing around those hooks.
Current boundary
The public repository is an unfinished prototype. The committed frontend references components that are not present, and the design document's activation patching and logit-lens goals are not implemented in the inspected API. It is source to examine, rather than a working browser demonstration.
Capturing an attention pattern is a starting point for inspection. It does not, by itself, explain a behavior or establish a causal mechanism. The next useful step is a bounded, reproducible intervention experiment supported by a complete interface.
Limitations
- Incomplete frontend; no working hosted demonstration is claimed.
- Activation patching and causal findings are not demonstrated by this repository.
- The underlying model instrumentation is provided by TransformerLens.
Source code & evidence
Results reflect the linked evidence snapshots. Source artifacts retain their dates and experimental conditions.