LLM explainer¶
Available
LLMExplainer is implemented and has a complete example with setup, executed outputs, and interpretation guidance.
The LLM explainer treats the language model as a black-box provider. It obtains the original response, perturbs the input prompt, measures semantic distance between that response and the perturbed prompt variants, and fits a local surrogate model to estimate term contributions to the resulting response-alignment score.
Current documented capabilities include:
- commercial, cloud, router, and local providers;
- runtime,
.env, and configuration-object setup; - text embeddings used by the semantic-distance workflow;
- automatic or selected surrogate models;
- token-level and contribution plots; and
- embedding comparison with fidelity metrics and executed outputs.
Open the LLM explainer API reference
Interpretation boundary¶
The result is a local perturbation-based response-alignment approximation. The current implementation queries the provider for the original response only; it does not reveal private chain-of-thought or reconstruct the model's exact internal computation.